Radiomics features on non-contrast-enhanced CT scan can precisely classify AVM-related hematomas from other spontaneous intraparenchymal hematoma types

Radiomics with Ultrasound Radiofrequency Data for Improving Evaluation of Duchenne Muscular Dystrophy

Duchenne muscular dystrophy (DMD) is a rare and severe genetic neuromuscular disease, characterized by rapid progression and high mortality, highlighting the need for accurate ambulatory function assessment tools. Ultrasound imaging methods have been widely used for quantitative analysis. Radiomics, which converts medical images into data, combined with machine learning (ML), offers a promising solution. This study is aimed at utilizing radiomics to analyze different stages of data generated during B-mode image processing to evaluate the ambulatory function of DMD patients. The study included 85 participants, categorized into ambulatory and non-ambulatory groups based on their functional status. Ultrasound scans were utilized to capture backscattered radiofrequency data, which were then processed to generate envelope, normalized, and B-mode images. Radiomics analysis involved the manual segmentation of grayscale images and automatic feature extraction using specialized software, followed by feature selection using the maximal relevance and minimal redundancy method. The selected features were input into five ML algorithms, with model evaluation conducted via area under the receiver operating characteristic curve (AUROC). To ensure robustness, both leave-one-out cross-validation and repeated data splitting methods were employed. Additionally, multiple ML models were constructed and tested to assess their performance. The intensity values across all image types increased as walking ability declined, with significant differences observed between the ambulatory and non-ambulatory groups (p < 0.001). These groups exhibited similar diagnostic performance levels, with AUROC values below 0.8. However, radiofrequency (RF) images outperformed other types when radiomics was applied, notably achieving an AUROC value of 0.906. Additionally, combining multiple ML algorithms yielded a higher AUROC value of 0.912 using RF images as input. Radiomics analysis of RF data surpasses conventional B-mode imaging and other ultrasound-derived images in evaluating ambulatory function in DMD. Moreover, integrating multiple machine learning models further enhances classification performance. The proposed method in this study offers a promising framework for improving the accuracy and reliability of clinical follow-up evaluations, supporting more effective management of DMD. The code is available at https://github.com/Goldenyan/radiomicsUS .

Machine learning based classification of spontaneous intracranial hemorrhages using radiomics features

PURPOSE

To assess the efficacy of radiomics features extracted from non-contrast computed tomography (NCCT) scans in differentiating multiple etiologies of spontaneous intracerebral hemorrhage (ICH).

METHODS

CT images and clinical data from 141 ICH patients from 2010 to 2022 were collected. The cohort comprised primary (n = 57), tumorous (n = 46), and vascular malformation-related ICH (n = 38). Radiomics features were extracted from the initial brain NCCT scans and identified potential features using mutual information. A hierarchical classification with AdaBoost classifiers was employed to classify the multiple etiologies of ICH. Age of the patient and ICH's location were examined alongside radiomics features. The accuracy, area under the curve (AUC), sensitivity, and specificity were used to evaluate classification performance.

RESULTS

The proposed method achieved an accuracy of 0.79. For identifying primary ICH, the model achieved a sensitivity of 0.86 and specificity of 0.87. Meanwhile, the sensitivity and specificity for identifying tumoral causes were 0.78 and 0.93, respectively. For vascular malformation, the model reached a sensitivity and specificity of 0.72 and 0.89, respectively. The AUCs for primary, tumorous, and vascular malformation were 0.86, 0.85, and 0.82, respectively. The findings further highlight the importance of texture-based variables in ICH classification. The age and location of the ICH can enhance the classification performance.

CONCLUSION

The use of a machine learning model with radiomics features has the potential in classifying the three types of non-traumatic ICH. It may help the radiologist decide on an appropriate further examination plan to arrive at a correct diagnosis.

Machine Learning Models for 3-Month Outcome Prediction Using Radiomics of Intracerebral Hemorrhage and Perihematomal Edema from Admission Head Computed Tomography (CT)

Background: Intracerebral hemorrhages (ICH) and perihematomal edema (PHE) are respective imaging markers of primary and secondary brain injury in hemorrhagic stroke. In this study, we explored the potential added value of PHE radiomic features for prognostication in ICH patients. Methods: Using a multicentric trial cohort of acute supratentorial ICH (n = 852) patients, we extracted radiomic features from ICH and PHE lesions on admission non-contrast head CTs. We trained and tested combinations of different machine learning classifiers and feature selection methods for prediction of poor outcome-defined by 4-to-6 modified Rankin Scale scores at 3-month follow-up-using five different input strategies: (a) ICH radiomics, (b) ICH and PHE radiomics, (c) admission clinical predictors of poor outcomes, (d) ICH radiomics and clinical variables, and (e) ICH and PHE radiomics with clinical variables. Models were trained on 500 patients, tested, and compared in 352 using the receiver operating characteristics Area Under the Curve (AUC), Integrated Discrimination Index (IDI), and Net Reclassification Index (NRI). Results: Comparing the best performing models in the independent test cohort, both IDI and NRI demonstrated better individual-level risk assessment by addition of PHE radiomics as input to ICH radiomics (both p < 0.001), but with insignificant improvement in outcome prediction (AUC of 0.74 versus 0.71, p = 0.157). The addition of ICH and PHE radiomics to clinical variables also improved IDI and NRI risk-classification (both p < 0.001), but with a insignificant increase in AUC of 0.85 versus 0.83 (p = 0.118), respectively. All machine learning models had greater or equal accuracy in outcome prediction compared to the widely used ICH score. Conclusions: The addition of PHE radiomics to hemorrhage lesion radiomics, as well as radiomics to clinical risk factors, can improve individual-level risk assessment, albeit with an insignificant increase in prognostic accuracy. Machine learning models offer quantitative and immediate risk stratification-on par with or more accurate than the ICH score-which can potentially guide patients' selection for interventions such as hematoma evacuation.

Role of [18F]-PSMA-1007 PET radiomics for seminal vesicle invasion prediction in primary prostate cancer

PURPOSE

The purpose of this study is to investigate the diagnostic utility of [18F]-PSMA-1007 PET radiomics combined with machine learning methods to predict seminal vesicle invasion (SVI) after radical prostatectomy (RP) in prostate cancer (PCa) patients.

METHODS

This is a post hoc retrospective analysis for a prospective clinical trial that included a consecutive sample of PCa patients (n = 140) who had [18F]-PSMA-1007 PET/CT prior to RP. The intraprostatic lesion's volume of interest (VOI) was semi-automatically sketched using a threshold of 40 % maximum standardized uptake value (SUVmax), namely 40%SUVmax-VOI, and seminal vesicle glands were manually contoured, namely SV-VOI. Models were built using a variety of machine learning methods such as logistic regression, random forest, and support vector machine. The area under the receiver operating characteristic curve (AUC) was calculated for different models, and the prediction performances of radiomics models were compared against the radiologists' assessment. Kaplan-Meier analysis was utilized to assess the effectiveness of selected radiomics features to determine the progression-free survival (PFS) probability.

RESULTS

The training set had 112 patients and the test set had 28 patients. The highest AUC for the PET radiomics model of 40%SUVmax-VOI and the PET radiomics model of SV-VOI were 0.85 and 0.96 in the test set, respectively. The PET radiomics model of SV-VOI had a significantly higher AUC compared to the radiologists' assessment (P < 0.05). The Kaplan-Meier analysis showed that PET radiomics features were associated with PFS in patients with PCa.

CONCLUSION

Radiomics models developed by preoperative [18F]-PSMA-1007 PET were proven useful in predicting SVI, and PSMA PET radiomics features were correlated with PFS, suggesting that the PSMA PET radiomics might be an accurate tool for PCa characterization.

Early prediction of acute gallstone pancreatitis severity: a novel machine learning model based on CT features and open access online prediction platform

BACKGROUND

Early diagnosis of acute gallstone pancreatitis severity (GSP) is challenging in clinical practice. We aimed to investigate the efficacy of CT features and radiomics for the early prediction of acute GSP severity.

METHODS

We retrospectively recruited GSP patients who underwent CT imaging within 48 h of admission from tertiary referral centre. Radiomics and CT features were extracted from CT scans. The clinical and CT features were selected by the random forest algorithm to develop the ML GSP model for the identification of severity of GSP (mild or severe), and its predictive efficacy was compared with radiomics model. The predictive performance was assessed by the area under operating characteristic curve. Calibration curve and decision curve analysis were performed to demonstrate the classification performance and clinical efficacy. Furthermore, we built a web-based open access GSP severity calculator. The study was registered with ClinicalTrials.gov (NCT05498961).

RESULTS

A total of 301 patients were enrolled. They were randomly assigned into the training (n = 210) and validation (n = 91) cohorts at a ratio of 7:3. The random forest algorithm identified the level of calcium ions, WBC count, urea level, combined cholecystitis, gallbladder wall thickening, gallstones, and hydrothorax as the seven predictive factors for severity of GSP. In the validation cohort, the areas under the curve for the radiomics model and ML GSP model were 0.841 (0.757-0.926) and 0.914 (0.851-0.978), respectively. The calibration plot shows that the ML GSP model has good consistency between the prediction probability and the observation probability. Decision curve analysis showed that the ML GSP model had high clinical utility.

CONCLUSIONS

We built the ML GSP model based on clinical and CT image features and distributed it as a free web-based calculator. Our results indicated that the ML GSP model is useful for predicting the severity of GSP.

Ischemic perfusion radiomics: assessing neurological impairment in acute ischemic stroke

Introduction

Accurate neurological impairment assessment is crucial for the clinical treatment and prognosis of patients with acute ischemic stroke (AIS). However, the original perfusion parameters lack the deep information for characterizing neurological impairment, leading to difficulty in accurate assessment. Given the advantages of radiomics technology in feature representation, this technology should provide more information for characterizing neurological impairment. Therefore, with its rigorous methodology, this study offers practical implications for clinical diagnosis by exploring the role of ischemic perfusion radiomics features in assessing the degree of neurological impairment.

Methods

This study employs a meticulous methodology, starting with generating perfusion parameter maps through Dynamic Susceptibility Contrast-Perfusion Weighted Imaging (DSC-PWI) and determining ischemic regions based on these maps and a set threshold. Radiomics features are then extracted from the ischemic regions, and the t-test and least absolute shrinkage and selection operator (Lasso) algorithms are used to select the relevant features. Finally, the selected radiomics features and machine learning techniques are used to assess the degree of neurological impairment in AIS patients.

Results

The results show that the proposed method outperforms the original perfusion parameters, radiomics features of the infarct and hypoxic regions, and their combinations, achieving an accuracy of 0.926, sensitivity of 0.923, specificity of 0.929, PPV of 0.923, NPV of 0.929, and AUC of 0.923, respectively.

Conclusion

The proposed method effectively assesses the degree of neurological impairment in AIS patients, providing an objective auxiliary assessment tool for clinical diagnosis.

The current landscape of machine learning-based radiomics in arteriovenous malformations: a systematic review and radiomics quality score assessment

Background

Arteriovenous malformations (AVMs) are rare vascular anomalies involving a disorganization of arteries and veins with no intervening capillaries. In the past 10 years, radiomics and machine learning (ML) models became increasingly popular for analyzing diagnostic medical images. The goal of this review was to provide a comprehensive summary of current radiomic models being employed for the diagnostic, therapeutic, prognostic, and predictive outcomes in AVM management.

Methods

A systematic literature review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, in which the PubMed and Embase databases were searched using the following terms: (cerebral OR brain OR intracranial OR central nervous system OR spine OR spinal) AND (AVM OR arteriovenous malformation OR arteriovenous malformations) AND (radiomics OR radiogenomics OR machine learning OR artificial intelligence OR deep learning OR computer-aided detection OR computer-aided prediction OR computer-aided treatment decision). A radiomics quality score (RQS) was calculated for all included studies.

Results

Thirteen studies were included, which were all retrospective in nature. Three studies (23%) dealt with AVM diagnosis and grading, 1 study (8%) gauged treatment response, 8 (62%) predicted outcomes, and the last one (8%) addressed prognosis. No radiomics model had undergone external validation. The mean RQS was 15.92 (range: 10-18).

Conclusion

We demonstrated that radiomics is currently being studied in different facets of AVM management. While not ready for clinical use, radiomics is a rapidly emerging field expected to play a significant future role in medical imaging. More prospective studies are warranted to determine the role of radiomics in the diagnosis, prediction of comorbidities, and treatment selection in AVM management.

Prediction of Obliteration After the Gamma Knife Radiosurgery of Arteriovenous Malformations Using Hand-Crafted Radiomics and Deep-Learning Methods

INTRODUCTION

Brain arteriovenous malformations (bAVMs) are vascular abnormalities that can be treated with embolization or radiotherapy to prevent the risk of future rupture. In this study, we use hand-crafted radiomics and deep learning techniques to predict favorable vs. unfavorable outcomes following Gamma Knife radiosurgery (GKRS) of bAVMs and compare their prediction performances.

METHODS

One hundred twenty-six patients seen at one academic medical center for GKRS obliteration of bAVMs over 15 years were retrospectively reviewed. Forty-two patients met the inclusion criteria. Favorable outcomes were defined as complete nidus obliteration demonstrated on cerebral angiogram and asymptomatic recovery. Unfavorable outcomes were defined as incomplete obliteration or complications relating to the AVM that developed after GKRS. Outcome predictions were made using a random forest model with hand-crafted radiomic features and a fine-tuned ResNet-34 convolutional neural network (CNN) model. The performance was evaluated by using a ten-fold cross-validation technique.

RESULTS

The average accuracy and area-under-curve (AUC) values of the Random Forest Classifier (RFC) with radiomics features were 68.5 ±9.80% and 0.705 ±0.086, whereas those of the ResNet-34 model were 60.0 ±11.9% and 0.694 ±0.124. Four radiomics features used with RFC discriminated unfavorable response cases from favorable response cases with statistical significance. When cropped images were used with ResNet-34, the accuracy and AUC decreased to 59.3 ± 14.2% and 55.4 ±10.4%, respectively.

CONCLUSIONS

A hand-crafted radiomics model and a pre-trained CNN model can be fine-tuned on pre-treatment MRI scans to predict clinical outcomes of AVM patients undergoing GKRS with equivalent prediction performance. The outcome predictions are promising but require further external validation on more patients.

Determining acute ischemic stroke onset time using machine learning and radiomics features of infarct lesions and whole brain

Accurate determination of the onset time in acute ischemic stroke (AIS) patients helps to formulate more beneficial treatment plans and plays a vital role in the recovery of patients. Considering that the whole brain may contain some critical information, we combined the Radiomics features of infarct lesions and whole brain to improve the prediction accuracy. First, the radiomics features of infarct lesions and whole brain were separately calculated using apparent diffusion coefficient (ADC), diffusion-weighted imaging (DWI) and fluid-attenuated inversion recovery (FLAIR) sequences of AIS patients with clear onset time. Then, the least absolute shrinkage and selection operator (Lasso) was used to select features. Four experimental groups were generated according to combination strategies: Features in infarct lesions (IL), features in whole brain (WB), direct combination of them (IW) and Lasso selection again after direct combination (IWS), which were used to evaluate the predictive performance. The results of ten-fold cross-validation showed that IWS achieved the best AUC of 0.904, which improved by 13.5% compared with IL (0.769), by 18.7% compared with WB (0.717) and 4.2% compared with IW (0.862). In conclusion, combining infarct lesions and whole brain features from multiple sequences can further improve the accuracy of AIS onset time.

Quantitative image signature and machine learning-based prediction of outcomes in cerebral cavernous malformations

PURPOSE

There is increasing interest in novel prognostic tools and predictive biomarkers to help identify, with more certainty, cerebral cavernous malformations (CCM) susceptible of bleeding if left untreated. We developed explainable quantitative-based machine learning models from magnetic resonance imaging (MRI) in a large CCM cohort to demonstrate the value of artificial intelligence and radiomics in complementing natural history studies for hemorrhage and functional outcome prediction.

MATERIALS AND METHODS

One-hundred-eighty-one patients from a prospectively registered cohort of 366 adults with CCM were included. Fluid attenuated inversion recovery (FLAIR) T2-weighted brain images were preprocessed, and CCM and surrounding edema were segmented before radiomic feature computation. Minority class oversampling, dimensionality reduction and feature selection methods were applied. With prospective hemorrhage as primary outcome, machine learning models were built, cross-validated, and compared using clinico-radiologic, radiomic, and combined features. SHapley Additive exPlanations (SHAP) was used for interpretation to determine the radiomic features with most contribution to hemorrhage prediction.

RESULTS

The highest performances in hemorrhage predictions on the test set were combining radiomic and clinico-radiological features with an area under the curve (AUC) of 83% using linear regression and selected features, and an F1 score of 61% and 85% sensitivity using K-nearest neighbors with principal component analysis (PCA). Multilayer perceptron had the best performance predicting modified Rankin Scale ≥ 2 with an AUC of 74% using PCA derived features. For interpretation of the selected radiomic signature XGBoost model, Shapley additive explanations highlighted 6 radiomic features contributing the most to hemorrhage prediction.

CONCLUSION

Quantitative image-based modeling using machine learning has the potential to highlight novel imaging biomarkers that predict hemorrhagic and functional outcomes, ensuring more precise and personalized care for CCM patients.

Detecting Double Expression Status in Primary Central Nervous System Lymphoma Using Multiparametric MRI Based Machine Learning

BACKGROUND

Double expression lymphoma (DEL) is a subtype of primary central nervous system lymphoma (PCNSL) that often has a poor prognosis. Currently, there are limited noninvasive ways to detect protein expression.

PURPOSE

To detect DEL in PCNSL using multiparametric MRI-based machine learning.

STUDY TYPE

Retrospective.

POPULATION

Forty PCNSL patients were enrolled in the study among whom 17 were DEL (9 males and 8 females, 61.29 ± 14.14 years) and 23 were non-DEL (14 males and 9 females, 55.57 ± 14.16 years) with 59 lesions (28 DEL and 31 non-DEL).

FIELD STRENGTH/SEQUENCE

ADC map derived from DWI (b = 0/1000 s/mm2 ), fast spin echo T2WI, T2FLAIR, and contrast-enhanced T1 weighted imaging (T1CE) were collected at 3.0 T.

ASSESSMENT

Two raters manually segmented lesions by ITK-SNAP on ADC, T2WI, T2FLAIR and T1CE. A total of 2234 radiomics features from the tumor segmentation area were extracted. The t-test was conducted to filter the features, and elastic net regression algorithm combined with recursive feature elimination was used to calculate the essential features. Finally, 12 groups with combinations of different sequences were fitted to 6 classifiers, and the optimal models were selected.

STATISTICAL TESTS

Continuous variables were assessed by the t-test, while categorical variables were assessed by the non-parametric test. Interclass correlation coefficient tested variables' consistency. Sensitivity, specificity, accuracy F1-score, and area under the curve (AUC) were used to evaluate model performance.

RESULTS

DEL status could be identified to varying degrees with 72 models based on radiomics, and model performance could be improved by combining different sequences and classifiers. Both SVMlinear and logistic regression (LR) combined with four sequence group had similar largest AUCmean (0.92 ± 0.09 vs. 0.92 ± 0.05), and SVMlinear was considered as the optimal model in this study since the F1-score of SVMlinear (0.88) was higher than that of LR (0.83).

DATA CONCLUSION

Multiparametric MRI-based machine learning is promising in DEL detection.

EVIDENCE LEVEL

4 TECHNICAL EFFICACY STAGE: 2.

High-resolution magnetic resonance imaging-based radiomic features aid in selecting endovascular candidates among patients with cerebral venous sinus thrombosis

OBJECTIVES

Cerebral venous sinus thrombosis (CVST) can cause sinus obstruction and stenosis, with potentially fatal consequences. High-resolution magnetic resonance imaging (HRMRI) can diagnose CVST qualitatively, although quantitative screening methods are lacking for patients refractory to anticoagulation therapy and who may benefit from endovascular treatment (EVT). Thus, in this study, we used radiomic features (RFs) extracted from HRMRI to build machine learning models to predict response to drug therapy and determine the appropriateness of EVT.

MATERIALS AND METHODS

RFs were extracted from three-dimensional T1-weighted motion-sensitized driven equilibrium (MSDE), T2-weighted MSDE, T1-contrast, and T1-contrast MSDE sequences to build radiomic signatures and support vector machine (SVM) models for predicting the efficacy of standard drug therapy and the necessity of EVT.

RESULTS

We retrospectively included 53 patients with CVST in a prospective cohort study, among whom 14 underwent EVT after standard drug therapy failed. Thirteen RFs were selected to construct the RF signature and CVST-SVM models. In the validation dataset, the sensitivity, specificity, and area under the curve performance for the RF signature model were 0.833, 0.937, and 0.977, respectively. The radiomic score was correlated with days from symptom onset, history of dyslipidemia, smoking, fibrin degradation product, and D-dimer levels. The sensitivity, specificity, and area under the curve for the CVST-SVM model in the validation set were 0.917, 0.969, and 0.992, respectively.

CONCLUSIONS

The CVST-SVM model trained with RFs extracted from HRMRI outperformed the RF signature model and could aid physicians in predicting patient responses to drug treatment and identifying those who may require EVT.

Radiomics features from 3D-MPRAGE imaging can differentiate temporal-plus epilepsy from temporal lobe epilepsy

OBJECTIVE

This study aimed to differentiate temporal-plus epilepsy (TPE) from temporal lobe epilepsy (TLE) using extraction of radiomics features from three-dimensional magnetization-prepared rapid acquisition gradient echo (3D-MPRAGE) imaging data.

METHODS

Data from patients with TLE or TPE who underwent epilepsy surgery between January 2019 and January 2021 were retrospectively analyzed. Thirty-three regions of interest in the affected hemisphere of each patient were defined on 3D-MPRAGE images. A total of 3531 image features were extracted from each patient. Four feature selection methods and 10 machine learning algorithms were used to build 40 differentiation models. Model performance was evaluated using receiver operating characteristic analysis.

RESULTS

Eighty-two patients were included for analysis, 47 with TLE and 35 with TPE. The model combining logistic regression and the relief selection method had the best performance (area under the receiver operating characteristic curve, .779; accuracy, .875; sensitivity, .800; specificity, .929; positive predictive value, .889; negative predictive value, .867).

SIGNIFICANCE

Radiomics analysis can differentiate TPE from TLE. The logistic regression classifier trained with radiomics features extracted from 3D-MPRAGE images had the highest accuracy and best performance.

Development and validation of a deep learning model for prediction of intracranial aneurysm rupture risk based on multi-omics factor

OBJECTIVE

The clinical ability of radiomics to predict intracranial aneurysm rupture risk remains unexplored. This study aims to investigate the potential uses of radiomics and explore whether deep learning (DL) algorithms outperform traditional statistical methods in predicting aneurysm rupture risk.

METHODS

This retrospective study included 1740 patients with 1809 intracranial aneurysms confirmed by digital subtraction angiography at two hospitals in China from January 2014 to December 2018. We randomly divided the dataset (hospital 1) into training (80%) and internal validation (20%). External validation was performed using independent data collected from hospital 2. The prediction models were developed based on clinical, aneurysm morphological, and radiomics parameters by logistic regression (LR). Additionally, the DL model for predicting aneurysm rupture risk using integration parameters was developed and compared with other models.

RESULTS

The AUCs of LR models A (clinical), B (morphological), and C (radiomics) were 0.678, 0.708, and 0.738, respectively (all p < 0.05). The AUCs of the combined feature models D (clinical and morphological), E (clinical and radiomics), and F (clinical, morphological, and radiomics) were 0.771, 0.839, and 0.849, respectively. The DL model (AUC = 0.929) outperformed the machine learning (ML) (AUC = 0.878) and the LR models (AUC = 0.849). Also, the DL model has shown good performance in the external validation datasets (AUC: 0.876 vs 0.842 vs 0.823, respectively).

CONCLUSION

Radiomics signatures play an important role in predicting aneurysm rupture risk. DL methods outperformed conventional statistical methods in prediction models for the rupture risk of unruptured intracranial aneurysms, integrating clinical, aneurysm morphological, and radiomics parameters.

KEY POINTS

• Radiomics parameters are associated with the rupture risk of intracranial aneurysms. • The prediction model based on integrating parameters in the deep learning model was significantly better than a conventional model. • The radiomics signature proposed in this study could guide clinicians in selecting appropriate patients for preventive treatment.

MRI-based machine learning models predict the malignant biological behavior of meningioma

BACKGROUND

The WHO grade and Ki-67 index are independent indices used to evaluate the malignant biological behavior of meningioma. This study aims to develop MRI-based machine learning models to predict the malignant biological behavior of meningioma from the perspective of the WHO grade, Ki-67 index, and their combination.

METHODS

This multicenter, retrospective study included 313 meningioma patients, of which 70 were classified as high-grade (WHO II/III) and 243 as low-grade (WHO I). The Ki-67 expression was classified into low-expression (n = 216) and high-expression (n = 97) groups with a threshold of 5%. Among them, there were 128 patients with malignant biological behavior whose WHO grade or Ki-67 index increased either or both. Data from Center A and B are were utilized for model development, while data from Center C and D were used for external validation. Radiomic features were extracted from the maximum cross-sectional area (2D) region of Interest (ROI) and the whole tumor volume (3D) ROI using different paraments from the T1, T2-weighted, and T1 contrast-enhanced sequences (T1CE), followed by five independent feature selections and eight classifiers. 240 prediction models were constructed to predict the WHO grade, Ki-67 index and their combination respectively. Models were evaluated by cross-validation in training set (n = 224). Suitable models were chosen by comparing the cross-validation (CV) area under the curves (AUC) and their relative standard deviations (RSD). Clinical and radiological features were collected and analyzed; meaningful features were combined with radiomic features to establish the clinical-radiological-radiomic (CRR) models. The receiver operating characteristic (ROC) analysis was used to evaluate those models in validation set. Radiomic models and CRR models were compared by Delong test.

RESULTS

1218 and 1781 radiomic features were extracted from 2D ROI and 3D ROI of each sequence. The selected grade, Ki-67 index and their combination radiomic models were T1CE-2D-LASSO-LR, T1CE-3D-LASSO-NB, and T1CE-2D-LASSO-LR, with cross-validated AUCs on the training set were 0.857, 0.798, and 0.888, the RSDs were 0.06, 0.09, and 0.05, the validation set AUCs were 0.829, 0.752, and 0.904, respectively. Heterogeneous enhancement was found to be associated with high grade and Ki-67 status, while surrounding invasion was associated with the high grade status, peritumoral edema and cerebrospinal fluid space surrounding tumor were correlated with the high Ki-67 status. The Delong test showed that these significant radiological features did not significantly improve the predictive performance. The AUCs for CRR models predicting grade, Ki-67 index, and their combination in the validation set were 0.821, 0.753, and 0.906, respectively.

CONCLUSIONS

This study demonstrated that MRI-based machine learning models could effectively predict the grade, Ki-67 index of meningioma. Models considering these two indices might be valuable for improving the predictive sensitivity and comprehensiveness of prediction of malignant biological behavior of meningioma.

Perihematomal edema-based CT-radiomics model to predict functional outcome in patients with intracerebral hemorrhage

PURPOSE

The purpose of this study was to identify possible association between noncontrast computed tomography (NCCT)-based radiomics features of perihematomal edema (PHE) and poor functional outcome at 90 days after intracerebral hemorrhage (ICH) and to develop a NCCT-based radiomics-clinical nomogram to predict 90-day functional outcomes in patients with ICH.

MATERIALS AND METHODS

In this multicenter retrospective study, 107 radiomics features were extracted from 1098 NCCT examinations obtained in 1098 patients with ICH. There were 652 men and 446 women with a mean age of 60 ± 12 (SD) years (range: 23-95 years). After harmonized and univariable and multivariable screening, seven of these radiomics features were closely associated with the 90-day functional outcome of patients with ICH. The radiomics score (Rad-score) was calculated based on the seven radiomics features. A clinical-radiomics nomogram was developed and validated in three cohorts. The model performance was evaluated using area under the curve analysis and decision and calibration curves.

RESULTS

Of the 1098 patients with ICH, 395 had a good outcome at 90 days. Hematoma hypodensity sign and intraventricular and subarachnoid hemorrhages were identified as risk factors for poor outcomes (P < 0.001). Age, Glasgow coma scale score, and Rad-score were independently associated with outcome. The clinical-radiomics nomogram showed good predictive performance with AUCs of 0.882 (95% CI: 0.859-0.905), 0.834 (95% CI: 0.776-0.891) and 0.905 (95% CI: 0.839-0.970) in the three cohorts and clinical applicability.

CONCLUSION

NCCT-based radiomics features from PHE are highly correlated with outcome. When combined with Rad-score, radiomics features from PHE can improve the predictive performance for 90-day poor outcome in patients with ICH.

MRI-based radiomics model for distinguishing Stage I endometrial carcinoma from endometrial polyp: a multicenter study

BACKGROUND

Patients with early endometrial carcinoma (EC) have a good prognosis, but it is difficult to distinguish from endometrial polyps (EPs).

PURPOSE

To develop and assess magnetic resonance imaging (MRI)-based radiomics models for discriminating Stage I EC from EP in a multicenter setting.

MATERIAL AND METHODS

Patients with Stage I EC (n = 202) and EP (n = 99) who underwent preoperative MRI scans were collected in three centers (seven devices). The images from devices 1-3 were utilized for training and validation, and the images from devices 4-7 were utilized for testing, leading to three models. They were evaluated by the area under the receiver operating characteristic curve (AUC) and metrics including accuracy, sensitivity, and specificity. Two radiologists evaluated the endometrial lesions and compared them with the three models.

RESULTS

The AUCs of device 1, 2_ada, device 1, 3_ada, and device 2, 3_ada for discriminating Stage I EC from EP were 0.951, 0.912, and 0.896 for the training set, 0.755, 0.928, and 1.000 for the validation set, and 0.883, 0.956, and 0.878 for the external validation set, respectively. The specificity of the three models was higher, but the accuracy and sensitivity were lower than those of radiologists.

CONCLUSION

Our MRI-based models showed good potential in differentiating Stage I EC from EP and had been validated in multiple centers. Their specificity was higher than that of radiologists and may be used for computer-aided diagnosis in the future to assist clinical diagnosis.

Machine learning-based radiomics analysis of preoperative functional liver reserve with MRI and CT image

OBJECTIVE

The indocyanine green retention rate at 15 min (ICG-R15) is a useful tool to evaluate the functional liver reserve before hepatectomy for liver cancer. Taking ICG-R15 as criteria, we investigated the ability of a machine learning (ML)-based radiomics model produced by Gd-EOB-DTPA-enhanced hepatic magnetic resonance imaging (MRI) or contrast-enhanced computed tomography (CT) image in evaluating functional liver reserve of hepatocellular carcinoma (HCC) patients.

METHODS

A total of 190 HCC patients with CT, among whom 112 also with MR, were retrospectively enrolled and randomly classified into a training dataset (CT: n = 133, MR: n = 78) and a test dataset (CT: n = 57, MR: n = 34). Then, radiomics features from Gd-EOB-DTPA MRI and CT images were extracted. The features associated with the ICG-R15 classification were selected. Five ML classifiers were used for the ML-model investigation. The accuracy (ACC) and the area under curve (AUC) of receiver operating characteristic (ROC) with 95% confidence intervals (CI) were utilized for ML-model performance evaluation.

RESULTS

A total of 107 different radiomics features were extracted from MRI and CT, respectively. The features related to ICG-R15 which was classified into 10%, 20% and 30% were selected. In MRI groups, classifier XGBoost performed best with its AUC = 0.917 and ACC = 0.882 when the threshold was set as ICG-R15 = 10%. When ICG-R15 = 20%, classifier Random Forest performed best with AUC = 0.979 and ACC = 0.882. When ICG-R15 = 30%, classifier XGBoost performed best with AUC = 0.961 and ACC = 0.941. For CT groups, the classifier XGBoost performed best when ICG-R15 = 10% with AUC = 0.822 and ACC = 0.842. When ICG-R15 = 20%, classifier SVM performed best with AUC = 0.860 and ACC = 0.842. When ICG-R15 = 30%, classifier XGBoost performed best with AUC = 0.938 and ACC = 0.965.

CONCLUSIONS

Both the MRI- and CT-based machine learning models are proved to be valuable noninvasive methods for functional liver reserve evaluation.

Combination of Hematoma Volume and Perihematoma Radiomics Analysis on Baseline CT Scan Predicts the Growth of Perihematomal Edema

PURPOSE

The aim is to explore the potential value of CT-based radiomics in predicting perihematomal edema (PHE) volumes after acute intracerebral hemorrhage (ICH) from admission to 24 h.

METHODS

A total of 231 patients newly diagnosed with acute ICH at two institutes were analyzed retrospectively. The patients were randomly divided into training (N = 117) and internal validation cohort (N = 45) from institute 1 with a ratio of 7:3. According to radiomics features extracted from baseline CT, the radiomics signatures were constructed. Multiple logistic regression analysis was used for clinical radiological factors and then the nomogram model was generated to predict the extent of PHE according to the optimal radiomics signature and the clinical radiological factors. The receiver operating characteristic (ROC) curve was used to evaluate the discrimination performance. The calibration curve and Hosmer-Lemeshow test were used to evaluate the consistency between the predicted and actual probability. The support vector regression (SVR) model was constructed to predict the overall value of follow-up PHE. The performance of the models was evaluated on the internal and independent validation cohorts.

RESULTS

The perihematoma 5 mm radiomics signature (AUC: 0.875) showed good ability to discriminate the small relative PHE(rPHE) from large rPHE volumes, comparing to intrahematoma radiomics signature (AUC: 0.711) or perihematoma 10 mm radiomics signature (AUC: 0.692) on the training cohort. The AUC of the combined nomogram model was 0.922 for the training cohort, 0.945 and 0.902 for the internal and independent validation cohorts, respectively. The calibration curves and Hosmer-Lemeshow test of the nomogram model suggested that the predictive performance and actual outcome were in favorable agreement. The SVR model also predicted the overall value of follow-up rPHE (root mean squared error, 0.60 and 0.45; Pearson correlation coefficient, 0.73 and 0.68; P < 0.001).

CONCLUSION

Among patients with acute ICH, the established nomogram and SVR model with favorable performance can offer a noninvasive tool for the prediction of PHE after ICH.

Development and validation of a clinical-radiomics model to predict recurrence for patients with hepatocellular carcinoma after curative resection

PURPOSE

Recurrence is the leading cause of death in hepatocellular carcinoma (HCC) patients with curative resection. In this study, we aimed to develop a preoperative predictive model based on high-throughput radiomics features and clinical factors for prediction of long- and short-term recurrence for these patients.

METHODS

A total of 270 patients with HCC who were followed up for at least 5 years after curative hepatectomy between June 2014 and December 2017 were enrolled in this retrospective study. Regions of interest were manually delineated in preoperative T2-weighted images using ITK-SNAP software on each HCC tumor slice. A total of 1197 radiomics features were extracted, and the recursive feature elimination method based on logistic regression was used for radiomics signature building. Tenfold cross-validation was applied for model development. Nomograms were constructed and assessed by calibration plot, which compares nomogram-predicated probability with observed outcome. Receiver-operating characteristic was then generated to evaluate the predictive performance of the model in the development and test cohorts.

RESULTS

The 10 most recurrence-free survival-related radiomics features were selected for the radiomics signatures. A multiparametric clinical-radiomics model combining albumin and radiomics score for recurrence prediction was further established. The integrated model demonstrated good calibration and satisfactory discrimination, with the area under the curve (AUC) of 0.864, 95% CI 0.842-0.903, sensitivity of 0.889, and specificity of 0.644 in the test set. Calibration curve showed good agreement concerning 5-year recurrence risk predicted by the nomogram. In addition, the AUC of 1-, 2-, 3-, and 4-year recurrence was 0.935 (95% CI 0.836-1.000), 0.861 (95% CI 0.723-0.999), 0.878 (95% CI 0.762-0.994), and 0.878 (95% CI 0.762-0.994) in the test set, respectively.

CONCLUSIONS

The clinical-radiomics model integrating radiomics features and clinical factors can improve recurrence predictions beyond predictions made using clinical factors or radiomics features alone. Our clinical-radiomics model is a valid method to predict recurrence that should improve preoperative prognostic performance and allow more individualized treatment decisions.

Introduction to radiomics for a clinical audience

Radiomics is a rapidly developing field of research focused on the extraction of quantitative features from medical images, thus converting these digital images into minable, high-dimensional data, which offer unique biological information that can enhance our understanding of disease processes and provide clinical decision support. To date, most radiomics research has been focused on oncological applications; however, it is increasingly being used in a raft of other diseases. This review gives an overview of radiomics for a clinical audience, including the radiomics pipeline and the common pitfalls associated with each stage. Key studies in oncology are presented with a focus on both those that use radiomics analysis alone and those that integrate its use with other multimodal data streams. Importantly, clinical applications outside oncology are also presented. Finally, we conclude by offering a vision for radiomics research in the future, including how it might impact our practice as radiologists.

A CT-based radiomics nomogram for classification of intraparenchymal hyperdense areas in patients with acute ischemic stroke following mechanical thrombectomy treatment

Objectives

To develop and validate a radiomic-based model for differentiating hemorrhage from iodinated contrast extravasation of intraparenchymal hyperdense areas (HDA) following mechanical thrombectomy treatment in acute ischemic stroke.

Methods

A total of 100 and four patients with intraparenchymal HDA on initial post-operative CT were included in this study. The patients who met criteria were divided into a primary and a validation cohort. A training cohort was constructed using Synthetic Minority Oversampling Technique on the primary cohort to achieve group balance. Thereafter, a radiomics score was calculated and the radiomic model was constructed. Clinical factors were assessed to build clinical model. Combined with the Rad-score and independent clinical factors, a combined model was constructed. Different models were assessed using the area under the receiver operator characteristic curves. The combined model was visualized as nomogram, and assessed with calibration and clinical usefulness.

Results

Cardiogenic diseases, intraoperative tirofiban administration and preoperative national institute of health stroke scale were selected as independent predictors to construct the clinical model with area under curve (AUC) of 0.756 and 0.693 in the training and validation cohort, respectively. Our data demonstrated that the radiomic model showed good discrimination in the training (AUC, 0.955) and validation cohort (AUC, 0.869). The combined nomogram model showed optimal discrimination in the training (AUC, 0.972) and validation cohort (AUC, 0.926). Decision curve analysis demonstrated the combined model had a higher overall net benefit in differentiating hemorrhage from iodinated contrast extravasation in terms of clinical usefulness.

Conclusions

The nomogram shows favorable efficacy for differentiating hemorrhage from iodinated contrast extravasation, which might provide an individualized tool for precision therapy.

Applicability of the Edinburgh CT Criteria for Lobar Intracerebral Hemorrhage Associated with Cerebral Amyloid Angiopathy

OBJECTIVE

Based on histopathology, Edinburgh diagnostic criteria were proposed to consider a nontraumatic intracerebral lobar hemorrhage (ICH) as related to cerebral amyloid angiopathy (CAA) using the initial computed tomography (CT) scan and the APOE genetic status. We aimed to externally validate the Edinburgh prediction model, excluding the APOE genotyping and based on the modified Boston criteria on the MRI for CAA diagnosis METHODS: We included patients admitted for spontaneous lobar ICH in the emergency department between 2016 and 2019 who underwent noncontrast CT scan and MRI. According to the MRI, patients were classified into the CAA group or into the non-CAA group in the case of other causes of ICH. Two neuroradiologists, blinded to the final retained diagnosis, rated each radiological feature on initial CT scan described in the Edinburgh study on initial CT scan RESULTS: A total of 102 patients were included, of whom 36 were classified in the CAA group, 46 in the non-CAA causes group and 20 of undetermined cause (excluded from the primary analysis). The Edinburgh prediction model, including finger-like projections and subarachnoid extension showed an area under receiver operating characteristic curves (AUC) of 0.760 (95% confidence interval, CI: 0.660-0.859) for the diagnosis of CAA. The AUC reached 0.808 (95% CI: 0.714-0.901) in a new prediction model integrating a third radiologic variable: the ICH cortical involvement.

CONCLUSION

Using the Boston MRI criteria as a final assessment, we provided a new external confirmation of the radiological Edinburgh CT criteria, which are directly applicable in acute settings of spontaneous lobar ICH and further proposed an original 3‑set model considering finger-like projections, subarachnoid extension, and cortical involvement that may achieve a high discrimination performance.

Computed tomography-based radiomics machine learning classifiers to differentiate type I and type II epithelial ovarian cancers

OBJECTIVES

To compare computed tomography (CT)-based radiomics for preoperatively differentiating type I and II epithelial ovarian cancers (EOCs) using different machine learning classifiers and to construct and validate the best diagnostic model.

METHODS

A total of 470 patients with EOCs were included retrospectively. Patients were divided into a training dataset (N = 329) and a test dataset (N = 141). A total of 1316 radiomics features were extracted from the portal venous phase of contrast-enhanced CT images for each patient, followed by dimension reduction of the features. The support vector machine (SVM), k-nearest neighbor (KNN), random forest (RF), naïve Bayes (NB), logistic regression (LR), and eXtreme Gradient Boosting (XGBoost) classifiers were trained to obtain the radiomics signatures. The performance of each radiomics signature was evaluated and compared by the area under the receiver operating characteristic curve (AUC) and relative standard deviation (RSD). The best radiomics signature was selected and combined with clinical and radiological features to establish a combined model. The diagnostic value of the combined model was assessed.

RESULTS

The LR-based radiomics signature performed well in the test dataset, with an AUC of 0.879 and an accuracy of 0.773. The combined model performed best in both the training and test datasets, with AUCs of 0.900 and 0.934 and accuracies of 0.848 and 0.823, respectively.

CONCLUSION

The combined model showed the best diagnostic performance for distinguishing between type I and II EOCs preoperatively. Therefore, it can be a useful tool for clinical individualized EOC classification.

KEY POINTS

• Radiomics features extracted from computed tomography (CT) could be used to differentiate type I and II epithelial ovarian cancers (EOCs). • Machine learning can improve the performance of differentiating type I and II EOCs. • The combined model exhibited the best diagnostic capability over the other models in both the training and test datasets.

The Combination of Whole-Brain Features and Local-Lesion Features in DSC-PWI May Improve Ischemic Stroke Outcome Prediction

Accurate and reliable outcome predictions can help evaluate the functional recovery of ischemic stroke patients and assist in making treatment plans. Given that recovery factors may be hidden in the whole-brain features, this study aims to validate the role of dynamic radiomics features (DRFs) in the whole brain, DRFs in local ischemic lesions, and their combination in predicting functional outcomes of ischemic stroke patients. First, the DRFs in the whole brain and the DRFs in local lesions of dynamic susceptibility contrast-enhanced perfusion-weighted imaging (DSC-PWI) images are calculated. Second, the least absolute shrinkage and selection operator (Lasso) is used to generate four groups of DRFs, including the outstanding DRFs in the whole brain (Lasso (WB)), the outstanding DRFs in local lesions (Lasso (LL)), the combination of them (combined DRFs), and the outstanding DRFs in the combined DRFs (Lasso (combined)). Then, the performance of the four groups of DRFs is evaluated to predict the functional recovery in three months. As a result, Lasso (combined) in the four groups achieves the best AUC score of 0.971, which improves the score by 8.9% compared with Lasso (WB), and by 3.5% compared with Lasso (WB) and combined DRFs. In conclusion, the outstanding combined DRFs generated from the outstanding DRFs in the whole brain and local lesions can predict functional outcomes in ischemic stroke patients better than the single DRFs in the whole brain or local lesions.

Radiomics features of DSC-PWI in time dimension may provide a new chance to identify ischemic stroke

Ischemic stroke has become a severe disease endangering human life. However, few studies have analyzed the radiomics features that are of great clinical significance for the diagnosis, treatment, and prognosis of patients with ischemic stroke. Due to sufficient cerebral blood flow information in dynamic susceptibility contrast perfusion-weighted imaging (DSC-PWI) images, this study aims to find the critical features hidden in DSC-PWI images to characterize hypoperfusion areas (HA) and normal areas (NA). This study retrospectively analyzed 80 DSC-PWI data of 56 patients with ischemic stroke from 2013 to 2016. For exploring features in HA and NA,13 feature sets (Fmethod) were obtained from different feature selection algorithms. Furthermore, these 13 Fmethod were validated in identifying HA and NA and distinguishing the proportion of ischemic lesions in brain tissue. In identifying HA and NA, the composite score (CS) of the 13 Fmethod ranged from 0.624 to 0.925. FLasso in the 13 Fmethod achieved the best performance with mAcc of 0.958, mPre of 0.96, mAuc of 0.982, mF1 of 0.959, and mRecall of 0.96. As to classifying the proportion of the ischemic region, the best CS was 0.786, with Acc of 0.888 and Pre of 0.863. The classification ability was relatively stable when the reference threshold (RT) was &lt;0.25. Otherwise, when RT was &gt;0.25, the performance will gradually decrease as its increases. These results showed that radiomics features extracted from the Lasso algorithms could accurately reflect cerebral blood flow changes and classify HA and NA. Besides, In the event of ischemic stroke, the ability of radiomics features to distinguish the proportion of ischemic areas needs to be improved. Further research should be conducted on feature engineering, model optimization, and the universality of the algorithms in the future.

Application of machine learning model to predict osteoporosis based on abdominal computed tomography images of the psoas muscle: a retrospective study

BACKGROUND

With rapid economic development, the world's average life expectancy is increasing, leading to the increasing prevalence of osteoporosis worldwide. However, due to the complexity and high cost of dual-energy x-ray absorptiometry (DXA) examination, DXA has not been widely used to diagnose osteoporosis. In addition, studies have shown that the psoas index measured at the third lumbar spine (L3) level is closely related to bone mineral density (BMD) and has an excellent predictive effect on osteoporosis. Therefore, this study developed a variety of machine learning (ML) models based on psoas muscle tissue at the L3 level of unenhanced abdominal computed tomography (CT) to predict osteoporosis.

METHODS

Medical professionals collected the CT images and the clinical characteristics data of patients over 40 years old who underwent DXA and abdominal CT examination in the Second Affiliated Hospital of Wenzhou Medical University database from January 2017 to January 2021. Using 3D Slicer software based on horizontal CT images of the L3, the specialist delineated three layers of the region of interest (ROI) along the bilateral psoas muscle edges. The PyRadiomics package in Python was used to extract the features of ROI. Then Mann-Whitney U test and the least absolute shrinkage and selection operator (LASSO) algorithm were used to reduce the dimension of the extracted features. Finally, six machine learning models, Gaussian naïve Bayes (GNB), random forest (RF), logistic regression (LR), support vector machines (SVM), Gradient boosting machine (GBM), and Extreme gradient boosting (XGBoost), were applied to train and validate these features to predict osteoporosis.

RESULTS

A total of 172 participants met the inclusion and exclusion criteria for the study. 82 participants were enrolled in the osteoporosis group, and 90 were in the non-osteoporosis group. Moreover, the two groups had no significant differences in age, BMI, sex, smoking, drinking, hypertension, and diabetes. Besides, 826 radiomic features were obtained from unenhanced abdominal CT images of osteoporotic and non-osteoporotic patients. Five hundred fifty radiomic features were screened out of 826 by the Mann-Whitney U test. Finally, 16 significant radiomic features were obtained by the LASSO algorithm. These 16 radiomic features were incorporated into six traditional machine learning models (GBM, GNB, LR, RF, SVM, and XGB). All six machine learning models could predict osteoporosis well in the validation set, with the area under the receiver operating characteristic (AUROC) values greater than or equal to 0.8. GBM is more effective in predicting osteoporosis, whose AUROC was 0.86, sensitivity 0.70, specificity 0.92, and accuracy 0.81 in validation sets.

CONCLUSION

We developed six machine learning models to predict osteoporosis based on psoas muscle images of abdominal CT, and the GBM model had the best predictive performance. GBM model can better help clinicians to diagnose osteoporosis and provide timely anti-osteoporosis treatment for patients. In the future, the research team will strive to include participants from multiple institutions to conduct external validation of the ML model of this study.

Assessment of gamma knife radiosurgery for unruptured cerebral arterioveneus malformations based on multi-parameter radiomics of MRI

OBJECTIVE

The treatment of Gamma knife radiosurgery (GKS) for unruptured Arteriovenous Malformations (AVM) remains controversial. A safe, effective and non-invasive method to predict outcome seems attractive for GKS. The purpose of this study was to develop and validate a MRI based multi-parameter radiomics model predicting the outcome of GKS for unruptured AVM.

METHODS

Eighty-eight unruptured AVM patients who initial underwent GKS between January 2011 and December 2016 in our hospital were included in this retrospective study. Patients were divided into two groups named as favourable and unfavourable outcome, according to the clinical outcome. Favourable outcome was defined as obliteration without post-SRS hemorrhage or permanent radiation-induced changes (RIC). Multivariate logistic regression analysis was used to select appropriate clinical features and construct a clinical predicting model. In terms of radiomic model, manually segmentation and radiomics extracted were performed on each AVM lesions. Finally, 1684 radiomics features were extracted and Recursive Feature Elimination (RFE) method combined with Random forest classifier were used for feature selection and model construction. The performance of the radiomics model was evaluated by the area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). In addition, the favourable group was further divided into early and late respond subgroup according to the time of obliteration evaluated by 2 years. The selected features were further compared according the respond time.

RESULTS

The median duration of neuroimaging follow-up was 65 months, 56 patients showed favourable outcome and 17 patients were observed obliteration within 2 years. The radiomics model constructed by 12 selected features achieved significant higher AUC of 0.88 (95% confidence interval 0.87-0.90) than traditional scoring system for predicting AVM outcome. Two selected radiomics features named "Dependence Variance" and "firstorder-Skewness" were found significant difference between the patients with early or late-respond.

CONCLUSIONS

The results suggest that the radiomics features could be successfully used for the pretreatment prediction of outcome for GKS in unruptured AVMs, which is helpful for decision-making process on unruptured AVM patients.

A Focus on the Role of DSC-PWI Dynamic Radiomics Features in Diagnosis and Outcome Prediction of Ischemic Stroke

Background: The ability to accurately detect ischemic stroke and predict its neurological recovery is of great clinical value. This study intended to evaluate the performance of whole-brain dynamic radiomics features (DRF) for ischemic stroke detection, neurological impairment assessment, and outcome prediction. Methods: The supervised feature selection (Lasso) and unsupervised feature-selection methods (five-feature dimension-reduction algorithms) were used to generate four experimental groups with DRF in different combinations. Ten machine learning models were used to evaluate their performance by ten-fold cross-validation. Results: In experimental group_A, the best AUCs (0.873 for stroke detection, 0.795 for NIHSS assessment, and 0.818 for outcome prediction) were obtained by outstanding DRF selected by Lasso, and the performance of significant DRF was better than the five-feature dimension-reduction algorithms. The selected outstanding dimension-reduction DRF in experimental group_C obtained a better AUC than dimension-reduction DRF in experimental group_A but were inferior to the outstanding DRF in experimental group_A. When combining the outstanding DRF with each dimension-reduction DRF (experimental group_B), the performance can be improved in ischemic stroke detection (best AUC = 0.899) and NIHSS assessment (best AUC = 0.835) but failed in outcome prediction (best AUC = 0.806). The performance can be further improved when combining outstanding DRF with outstanding dimension-reduction DRF (experimental group_D), achieving the highest AUC scores in all three evaluation items (0.925 for stroke detection, 0.853 for NIHSS assessment, and 0.828 for outcome prediction). By the method in this study, comparing the best AUC of Ft-test in experimental group_A and the best_AUC in experimental group_D, the AUC in stroke detection increased by 19.4% (from 0.731 to 0.925), the AUC in NIHSS assessment increased by 20.1% (from 0.652 to 0.853), and the AUC in prognosis prediction increased by 14.9% (from 0.679 to 0.828). This study provided a potential clinical tool for detailed clinical diagnosis and outcome prediction before treatment.

Machine-learning-based contrast-enhanced computed tomography radiomic analysis for categorization of ovarian tumors

Objectives

This study aims to evaluate the diagnostic performance of machine-learning-based contrast-enhanced CT radiomic analysis for categorizing benign and malignant ovarian tumors.

Methods

A total of 1,329 patients with ovarian tumors were randomly divided into a training cohort (N=930) and a validation cohort (N=399). All tumors were resected, and pathological findings were confirmed. Radiomic features were extracted from the portal venous phase images of contrast-enhanced CT. The clinical predictors included age, CA-125, HE-4, ascites, and margin of tumor. Both radiomics model (including selected radiomic features) and mixed model (incorporating selected radiomic features and clinical predictors) were constructed respectively. Six classifiers [k-nearest neighbor (KNN), support vector machines (SVM), random forest (RF), logistic regression (LR), multi-layer perceptron (MLP), and eXtreme Gradient Boosting (XGBoost)] were used for each model. The mean relative standard deviation (RSD) and area under the receiver operating characteristic curve (AUC) were applied to evaluate and select the best classifiers. Then, the performances of the two models with selected classifiers were assessed in the validation cohort.

Results

The MLP classifier with the least RSD (1.21 and 0.53, respectively) was selected as the best classifier in both radiomics and mixed models. The two models with MLP classifier performed well in the validation cohort, with the AUCs of 0.91 and 0.96 and with accuracies (ACCs) of 0.83 and 0.87, respectively. The Delong test showed that the AUC of mixed model was statistically different from that of radiomics model (p&lt;0.001).

Conclusions

Machine-learning-based CT radiomic analysis could categorize ovarian tumors with good performance preoperatively. The mixed model with MLP classifier may be a potential tool in clinical applications.

Novel Survival Features Generated by Clinical Text Information and Radiomics Features May Improve the Prediction of Ischemic Stroke Outcome

BACKGROUND

Accurate outcome prediction is of great clinical significance in customizing personalized treatment plans, reducing the situation of poor recovery, and objectively and accurately evaluating the treatment effect. This study intended to evaluate the performance of clinical text information (CTI), radiomics features, and survival features (SurvF) for predicting functional outcomes of patients with ischemic stroke.

METHODS

SurvF was constructed based on CTI and mRS radiomics features (mRSRF) to improve the prediction of the functional outcome in 3 months (90-day mRS). Ten machine learning models predicted functional outcomes in three situations (2-category, 4-category, and 7-category) using seven feature groups constructed by CTI, mRSRF, and SurvF.

RESULTS

For 2-category, ALL (CTI + mRSRF+ SurvF) performed best, with an mAUC of 0.884, mAcc of 0.864, mPre of 0.877, mF1 of 0.86, and mRecall of 0.864. For 4-category, ALL also achieved the best mAuc of 0.787, while CTI + SurvF achieved the best score with mAcc = 0.611, mPre = 0.622, mF1 = 0.595, and mRe-call = 0.611. For 7-category, CTI + SurvF performed best, with an mAuc of 0.788, mPre of 0.519, mAcc of 0.529, mF1 of 0.495, and mRecall of 0.47.

CONCLUSIONS

The above results indicate that mRSRF + CTI can accurately predict functional outcomes in ischemic stroke patients with proper machine learning models. Moreover, combining SurvF will improve the prediction effect compared with the original features. However, limited by the small sample size, further validation on larger and more varied datasets is necessary.

A Radiomics Nomogram for Classifying Hematoma Entities in Acute Spontaneous Intracerebral Hemorrhage on Non-contrast-Enhanced Computed Tomography

Aim

To develop and validate a radiomics nomogram on non-contrast-enhanced computed tomography (NECT) for classifying hematoma entities in patients with acute spontaneous intracerebral hemorrhage (ICH).

Materials and Methods

One hundred and thirty-five patients with acute intraparenchymal hematomas and baseline NECT scans were retrospectively analyzed, i.e., 52 patients with vascular malformation-related hemorrhage (VMH) and 83 patients with primary intracerebral hemorrhage (PICH). The patients were divided into training and validation cohorts in a 7:3 ratio with a random seed. After extracting the radiomics features of hematomas from baseline NECT, the least absolute shrinkage and selection operator (LASSO) regression was applied to select features and construct the radiomics signature. Multivariate logistic regression analysis was used to determine the independent clinical-radiological risk factors, and a clinical model was constructed. A predictive radiomics nomogram was generated by incorporating radiomics signature and clinical-radiological risk factors. Nomogram performance was assessed in the training cohort and tested in the validation cohort. The capability of models was compared by calibration, discrimination, and clinical benefit.

Results

Six features were selected to establish radiomics signature via LASSO regression. The clinical model was constructed with the combination of age [odds ratio (OR): 6.731; 95% confidence interval (CI): 2.209–20.508] and hemorrhage location (OR: 0.089; 95% CI: 0.028–0.281). Radiomics nomogram [area under the curve (AUC), 0.912 and 0.919] that incorporated age, location, and radiomics signature outperformed the clinical model (AUC, 0.816 and 0.779) and signature (AUC, 0.857 and 0.810) in the training cohort and validation cohorts, respectively. Good calibration and clinical benefit of nomogram were achieved in the training and validation cohorts.

Conclusion

Non-contrast-enhanced computed tomography-based radiomics nomogram can predict the individualized risk of VMH in patients with acute ICH.

Angioarchitectural features of brain arteriovenous malformation presented with seizures

Seizures are the second most common manifestations of brain arteriovenous malformations (bAVMs). This study was conducted to investigate the clinical and angioarchitectural features of bAVMs with seizures and provide guidelines for the clinical management of these patients. We collected clinical and radiological data on patients with bAVMs diagnosed by digital subtraction angiography between January 2013 and December 2020 and dichotomized the patients into the seizures and non-seizures groups. We identified differences in demographic and angiographic features. Logistic regression and random forest (RF) models were developed and compared. The diagnostic capacity was assessed using receiver operating characteristic (ROC) curves. A nomogram was constructed, and the clinical impact was determined by decision curve analysis. A total of 414 patients with bAVMs were included in the analysis, of which 78 (18.8%) had bAVM-related seizures. In the multivariable logistic regression model, the location and side of bAVMs were independently associated with seizures. In RF models, the maximal diameter of veins and the cross-sectional area of feeding arteries and draining veins were the most important features. ROC curves showed that the RF model was not better than MLR in predicting seizures. Decision curve analysis revealed that the use of a constructed nomogram to stratify the seizure patients was beneficial at all threshold probabilities in our study. The side and location of bAVMs are specific angioarchitectural features independently associated with the occurrences of seizures with bAVMs.

Machine Learning-Based Radiomics for Prediction of Epidermal Growth Factor Receptor Mutations in Lung Adenocarcinoma

Identifying an epidermal growth factor receptor (EGFR) mutation is important because EGFR tyrosine kinase inhibitors are the first-line treatment of choice for patients with EGFR mutation-positive lung adenocarcinomas (LUAC). This study is aimed at developing and validating a radiomics-based machine learning (ML) approach to identify EGFR mutations in patients with LUAC. We retrospectively collected data from 201 patients with positive EGFR mutation LUAC (140 in the training cohort and 61 in the validation cohort). We extracted 1316 radiomics features from preprocessed CT images and selected 14 radiomics features and 1 clinical feature which were most relevant to mutations through filter method. Subsequently, we built models using 7 ML approaches and established the receiver operating characteristic (ROC) curve to assess the discriminating performance of these models. In terms of predicting EGFR mutation, the model derived from radiomics features and combined models (radiomics features and relevant clinical factors) had an AUC of 0.79 (95% confidence interval (CI): 0.77-0.82), 0.86 (0.87-0.88), respectively. Our study offers a radiomics-based ML model using filter methods to detect the EGFR mutation in patients with LUAC. This convenient and low-cost method may be of help to noninvasively identify patients before obtaining tumor sample for molecule testing.

A Machine Learning Model Predicts the Outcome of SRS for Residual Arteriovenous malformations after partial embolization- A Real-World Clinical Obstacle

PURPOSE

To propose a machine learning (ML) model predicting the favorable outcome of stereotactic radiosurgery (SRS) for residual brain arteriovenous malformation (bAVM) after partial embolization.

MATERIALS AND METHODS

One hundred and thirty bAVM patients who underwent partial embolization followed by SRS were retrospectively reviewed. Patients were randomly split into training datasets (n=100) and testing datasets (n=30). Radiomics and dosimetric features were extracted from pre-SRS treatment images. Feature selection was performed to select appropriate radiomics and dosimetric features. Three ML algorithms were applied to construct models using selected features respectively. A total of 9 models were trained to predict favorable outcomes (obliteration without complication) of bAVMs. The efficacy of these models was evaluated on the testing dataset using mean accuracy (ACC) and area under the receiver operating characteristic curve (AUROC).

RESULTS

The obliteration rate of this cohort was 70.77% (92/130) with a mean follow-up period of 43.8 (Range 12-108 months) months. Favorable outcomes were achieved in 89 (68.46%) patients. Four radiomics features and 7 dosimetric features were selected for ML model construction. The dosimetric SVM showed the best performance on the training dataset, with an ACC and AUC of 0.74 and 0.78 respectively. The dosimetric SVM model also showed the best performance on the testing dataset where the ACC and AUC were 0.83 and 0.77 respectively.

CONCLUSION

Dosimetric features are good predictors of prognosis for patients with partially embolized bAVM followed by SRS therapy. The use of ML models is an innovative method for predicting favorable outcomes of partially embolized bAVM followed by SRS therapy.

Optimal machine learning methods for radiomic prediction models: Clinical application for preoperative T2*-weighted images of cervical spondylotic myelopathy

INTRODUCTION

Predicting the postoperative neurological function of cervical spondylotic myelopathy (CSM) patients is generally based on conventional magnetic resonance imaging (MRI) patterns, but this approach is not completely satisfactory. This study utilized radiomics, which produced advanced objective and quantitative indicators, and machine learning to develop, validate, test, and compare models for predicting the postoperative prognosis of CSM.

MATERIALS AND METHODS

In total, 151 CSM patients undergoing surgical treatment and preoperative MRI was retrospectively collected and divided into good/poor outcome groups based on postoperative modified Japanese Orthopedic Association (mJOA) scores. The datasets obtained from several scanners (an independent  scanner) for the training (testing) cohort were used for cross-validation (CV). Radiological models based on the intramedullary hyperintensity and compression ratio were constructed with 14 binary classifiers. Radiomic models based on 237 robust radiomic features were constructed with the same 14 binary classifiers in combination with 7 feature reduction methods, resulting in 98 models. The main outcome measures were the area under the receiver operating characteristic curve (AUROC) and accuracy.

RESULTS

Forty-one (11) radiomic models were superior to random guessing during CV (testing), with significant increased AUROC and/or accuracy (P AUROC < .05 and/or P accuracy < .05). One radiological model performed better than random guessing during CV (P accuracy < .05). In the testing cohort, the linear SVM preprocessor + SVM, the best radiomic model (AUROC: 0.74 ± 0.08, accuracy: 0.73 ± 0.07), overperformed the best radiological model (P AUROC = .048).

CONCLUSION

Radiomic features can predict postoperative spinal cord function in CSM patients. The linear SVM preprocessor + SVM has great application potential in building radiomic models.

Machine learning-based Radiomics analysis for differentiation degree and lymphatic node metastasis of extrahepatic cholangiocarcinoma

Background

Radiomics may provide more objective and accurate predictions for extrahepatic cholangiocarcinoma (ECC). In this study, we developed radiomics models based on magnetic resonance imaging (MRI) and machine learning to preoperatively predict differentiation degree (DD) and lymph node metastasis (LNM) of ECC.

Methods

A group of 100 patients diagnosed with ECC was included. The ECC status of all patients was confirmed by pathology. A total of 1200 radiomics features were extracted from axial T1 weighted imaging (T1WI), T2-weighted imaging (T2WI), diffusion weighted imaging (DWI), and apparent diffusion coefficient (ADC) images. A systematical framework considering combinations of five feature selection methods and ten machine learning classification algorithms (classifiers) was developed and investigated. The predictive capabilities for DD and LNM were evaluated in terms of area under precision recall curve (AUPRC), area under the receiver operating characteristic (ROC) curve (AUC), negative predictive value (NPV), accuracy (ACC), sensitivity, and specificity. The prediction performance among models was statistically compared using DeLong test.

Results

For DD prediction, the feature selection method joint mutual information (JMI) and Bagging Classifier achieved the best performance (AUPRC = 0.65, AUC = 0.90 (95% CI 0.75–1.00), ACC = 0.85 (95% CI 0.69–1.00), sensitivity = 0.75 (95% CI 0.30–0.95), and specificity = 0.88 (95% CI 0.64–0.97)), and the radiomics signature was composed of 5 selected features. For LNM prediction, the feature selection method minimum redundancy maximum relevance and classifier eXtreme Gradient Boosting achieved the best performance (AUPRC = 0.95, AUC = 0.98 (95% CI 0.94–1.00), ACC = 0.90 (95% CI 0.77–1.00), sensitivity = 0.75 (95% CI 0.30–0.95), and specificity = 0.94 (95% CI 0.72–0.99)), and the radiomics signature was composed of 30 selected features. However, these two chosen models were not significantly different to other models of higher AUC values in DeLong test, though they were significantly different to most of all models.

Conclusion

MRI radiomics analysis based on machine learning demonstrated good predictive accuracies for DD and LNM of ECC. This shed new light on the noninvasive diagnosis of ECC.

Artificial Intelligence Technologies in Neurosurgery: a Systematic Literature Review Using Topic Modeling. Part II: Research Objectives and Perspectives

The current increase in the number of publications on the use of artificial intelligence (AI) technologies in neurosurgery indicates a new trend in clinical neuroscience.

The aim of the study was to conduct a systematic literature review to highlight the main directions and trends in the use of AI in neurosurgery.

Temporal-Spatial Feature Extraction of DSA Video and Its Application in AVM Diagnosis

Objectives: Brain arteriovenous malformation (AVM) is one of the most common causes of intracranial hemorrhage in young adults, and its expeditious diagnosis on digital subtraction angiography (DSA) is essential for clinical decision-making. This paper firstly proposed a deep learning network to extract vascular time-domain features from DSA videos. Then, the temporal features were combined with spatial radiomics features to build an AVM-assisted diagnosis model. Materials and method: Anteroposterior position (AP) DSA videos from 305 patients, 153 normal and 152 with AVM, were analyzed. A deep learning network based on Faster-RCNN was proposed to track important vascular features in DSA. Then the appearance order of important vascular structures was quantified as the temporal features. The structure distribution and morphological features of vessels were quantified as 1,750 radiomics features. Temporal features and radiomics features were fused in a classifier based on sparse representation and support vector machine. An AVM diagnosis and grading system that combined the temporal and spatial radiomics features of DSA was finally proposed. Accuracy (ACC), sensitivity (SENS), specificity (SPEC), and area under the receiver operating characteristic curve (AUC) were calculated to evaluate the performance of the radiomics model. Results: For cerebrovascular structure detection, the average precision (AP) was 0.922, 0.991, 0.769, 0.899, and 0.929 for internal carotid artery, Willis circle, vessels, large veins, and venous sinuses, respectively. The mean average precision (mAP) of five time phases was 0.902. For AVM diagnosis, the models based on temporal features, radiomics features, and combined features achieved AUC of 0.916, 0.918, and 0.942, respectively. In the AVM grading task, the proposed combined model also achieved AUC of 0.871 in the independent testing set. Conclusion: DSA videos provide rich temporal and spatial distribution characteristics of cerebral blood vessels. Clinicians often interpret these features based on subjective experience. This paper proposes a scheme based on deep learning and traditional machine learning, which effectively integrates the complex spatiotemporal features in DSA, and verifies the value of this scheme in the diagnosis of AVM.

Can perihaematomal radiomics features predict haematoma expansion?

AIM

To evaluate the association between perihaematomal radiomics features and haematoma expansion (HE).

MATERIALS AND METHODS

Clinical and radiological data were collected retrospectively. The 1:1 propensity score matching (PSM) method was used to balance the difference of baseline characteristics between patients with and without HE. Radiomics features were extracted from the intra- and perihaematomal regions. Top HE-associated features were selected using the minimum redundancy, maximum relevancy algorithm. Support vector machine models were used to predict HE. Predictive performance of radiomics features from different regions was evaluated by receiver operating characteristic curve and confusion matrix-derived metrics.

RESULTS

A total of 1,062 patients were enrolled. After PSM analysis, the propensity score-matched cohort (PSM cohort) included 314 patients (HE: n=157; non-HE: n=157). The PSM cohort was distributed into the training (n=218) and the validation cohorts (n=96). The predictive performance of intra- and perihaematomal features were comparable in the training (area under the receiver operating characteristic curve [AUC], 0.751 versus 0.757; p=0.867) and the validation cohorts (AUC, 0.724 versus 0.671; p=0.454). By incorporating intra- and perihaematomal features, the combined model outperformed the single intrahaematomal model in the training cohort (AUC, 0.872 versus 0.751; p<0.001). Decision curve analysis (DCA) further confirmed the clinical usefulness of the combined model.

CONCLUSION

Perihaematomal radiomics features can predict HE. The integration of intra- and perihaematomal signatures may provide additional benefit to the prediction of HE.

IL-33 as a Novel Serum Prognostic Marker of Intracerebral Hemorrhage

Objective

Interleukin 33 (IL-33) is a key cytokine involved in inflammation and oxidative stress. The significance of serum IL-33 levels on the prognosis of patients with intracerebral hemorrhage (ICH) has not been well studied. The purpose of this study is to determine whether there is a relationship between the serum IL-33 level and the prognosis of patients with ICH upon admission.

Methods

A total of 402 patients with confirmed ICH were included in this study. Their demographic data, medical history, laboratory data, imaging data, and clinical scores on admission were collected. At the same time, enzyme-linked immunoassay (ELISA) was used to detect the serum IL-33 levels of patients. The prognosis of patients was evaluated by mRS scale after 3 months, and mRS > 2 was defined as poor prognosis.

Results

Among 402 patients with ICH, the number of patients with good prognosis and poor prognosis after 3 months was 148 and 254, respectively. Compared with the ICH group with poor prognosis, the ICH group with good prognosis had lower baseline NHISS scores (p = 0.039) and hematoma volume (p = 0.025) and higher GCS scores (p < 0.001) and serum IL-33 levels (p < 0.001). The results of linear correlation analysis showed that serum IL-33 levels were significantly negatively correlated with baseline NHISS scores (r = −0.224, p = 0.033) and hematoma volume (r = −0.253, p = 0.046) but were significantly positively correlated with baseline GCS scores (r = 0.296, p = 0.020). The receiver operating characteristic curve (ROC) analysis showed that the sensitivity and specificity of serum IL-33 level in evaluating the prognosis of ICH were 72.1% and 74.3%, respectively. A cut-off value of serum IL-33 level < 109.3 pg/mL may indicate a poor prognosis for ICH.

Conclusions

Serum IL-33 level on admission may be a prognostic indicator of ICH, and its underlying mechanism needs further study.

Clinical-radiomics Nomogram for Risk Estimation of Early Hematoma Expansion after Acute Intracerebral Hemorrhage

RATIONALE AND OBJECTIVES

Noncontrast CT-based radiomics signature has shown ability for detecting hematoma expansion (HE) in spontaneous intracerebral hemorrhage (ICH). We sought to compare its predictive performance with clinical risk factors and develop a clinical-radiomics nomogram to assess the risk of early HE.

MATERIALS AND METHODS

In total, 1153 patients with ICH who underwent baseline cranial CT within 6 hours and follow-up scans within 72 hours of stroke onset were enrolled, of whom 864 (75%) were assigned to the derivation cohort and 289 (25%) to the validation cohort. Based on LASSO algorithm or stepwise logistic regression analysis, three models (clinical model, radiomics model, and hybrid model) were constructed to predict HE. The Akaike information criterion (AIC) and likelihood ratio test (LRT) were used for comparing the goodness of fit of the three models, and the AUC was used to evaluate their discrimination ability for HE.

RESULTS

The hybrid model (AIC = 681.426; χ²= 128.779) was the optimal model with the lowest AIC and highest chi-square values compared to the radiomics model (AIC = 767.979; χ² = 110.234) or the clinical model (AIC = 753.757; χ² = 56.448). The radiomics model was superior in the prediction of HE to the clinical model in both derivation (p = 0.009) and validation (p = 0.022) cohorts. In both datasets, the clinical-radiomics nomogram showed satisfactory discrimination and calibration for detecting HE (AUC = 0.771, Sensitivity = 87.0%; AUC = 0.820, Sensitivity = 88.1%; respectively).

CONCLUSION

Among patients with acute ICH, noncontrast CT-based radiomics model outperformed the clinical-only model in the prediction of HE, and the established clinical-radiomics nomogram with favorable performance can offer a noninvasive tool for the risk stratification of HE.

Radiomics in Stroke Neuroimaging: Techniques, Applications, and Challenges

Stroke is a leading cause of disability and mortality worldwide, resulting in substantial economic costs for post-stroke care each year. Neuroimaging, such as cranial computed tomography or magnetic resonance imaging, is the backbone of stroke management strategies, which can guide treatment decision-making (thrombolysis or hemostasis) at an early stage. With advances in computational technologies, particularly in machine learning, visual image information can now be converted into numerous quantitative features in an objective, repeatable, and high-throughput manner, in a process known as radiomics. Radiomics is mainly used in the field of oncology, which remains an area of active research. Over the past few years, investigators have attempted to apply radiomics to stroke in the hope of gaining benefits similar to those obtained in cancer management, i.e., in promoting the development of personalized precision medicine. Currently, radiomic analysis has shown promise for a variety of applications in stroke, including the diagnosis of stroke lesions, early prediction of outcomes, and evaluation for long-term prognosis. In this article, we elaborate the contributions of radiomics to stroke, as well as the subprocesses and techniques involved in radiomics studies. We also discuss the potential challenges facing its widespread implementation in routine practice and the directions for future research.

Artificial Intelligence Applications for Workflow, Process Optimization and Predictive Analytics

There is great potential for artificial intelligence (AI) applications, especially machine learning and natural language processing, in medical imaging. Much attention has been garnered by the image analysis tasks for diagnostic decision support and precision medicine, but there are many other potential applications of AI in radiology and have potential to enhance all levels of the radiology workflow and practice, including workflow optimization and support for interpretation tasks, quality and safety, and operational efficiency. This article reviews the important potential applications of informatics and AI related to process improvement and operations in the radiology department.

The first use of artificial intelligence (AI) in the ER: triage not diagnosis

Predictions related to the impact of AI on radiology as a profession run the gamut from AI putting radiologists out of business to having no effect at all. The use of AI appears to show significant promise in ER triage in the present. We briefly discuss the emerging effectiveness of AI in the ER imaging setting by looking at some of the products approved by the FDA and finding their way into "practice." The FDA approval process to date has focused on applications that affect patient triage and not necessarily ones that have the computer serve as the only or final reader. We describe a select group of applications to provide the reader with a sense of the current state of AI use in the ER setting to assess neurologic, pulmonary, and musculoskeletal trauma indications. In the process, we highlight the benefits of triage staging using AI, such as accelerating diagnosis and optimizing workflow, with few downsides. The ability to triage patients and take care of acute processes such as intracranial bleed, pneumothorax, and pulmonary embolism will largely benefit the health system, improving patient care and reducing costs. These capabilities are all available now. This first wave of AI applications is not replacing radiologists. Rather, the innovative software is improving throughput, contributing to the timeliness in which radiologists can get to read abnormal scans, and possibly enhances radiologists' accuracy. As for what the future holds for the use of AI in radiology, only time will tell.

Differentiation of supratentorial single brain metastasis and glioblastoma by using peri-enhancing oedema region-derived radiomic features and multiple classifiers

OBJECTIVE

To differentiate supratentorial single brain metastasis (MET) from glioblastoma (GBM) by using radiomic features derived from the peri-enhancing oedema region and multiple classifiers.

METHODS

One hundred and twenty single brain METs and GBMs were retrospectively reviewed and then randomly divided into a training data set (70%) and validation data set (30%). Quantitative radiomic features of each case were extracted from the peri-enhancing oedema region of conventional MR images. After feature selection, five classifiers were built. Additionally, the combined use of the classifiers was studied. Accuracy, sensitivity, and specificity were used to evaluate the classification performance.

RESULTS

A total of 321 features were extracted, and 3 features were selected for each case. The 5 classifiers showed an accuracy of 0.70 to 0.76, sensitivity of 0.57 to 0.98, and specificity of 0.43 to 0.93 for the training data set, with an accuracy of 0.56 to 0.64, sensitivity of 0.39 to 0.78, and specificity of 0.50 to 0.89 for the validation data set. When combining the classifiers, the classification performance differed according to the combined mode and the agreement pattern of classifiers, and the greatest benefit was obtained when all the classifiers reached agreement using the same weight and simple majority vote method.

CONCLUSIONS

Three features derived from the peri-enhancing oedema region had moderate value in differentiating supratentorial single brain MET from GBM with five single classifiers. Combined use of classifiers, like multi-disciplinary team (MDT) consultation, could confer extra benefits, especially for those cases when all classifiers reach agreement.

KEY POINTS

• Radiomics provides a way to differentiate single brain MET between GBM by using conventional MR images. • The results of classifiers or algorithms themselves are also data, the transformation of the primary data. • Like MDT consultation, the combined use of multiple classifiers may confer extra benefits.

Prediction of pulmonary pressure after Glenn shunts by computed tomography-based machine learning models

OBJECTIVES

This study aimed to develop non-invasive machine learning classifiers for predicting post-Glenn shunt patients with low and high risks of a mean pulmonary arterial pressure (mPAP) > 15 mmHg based on preoperative cardiac computed tomography (CT).

METHODS

This retrospective study included 96 patients with functional single ventricle who underwent a bidirectional Glenn procedure between November 1, 2009, and July, 31, 2017. All patients underwent post-procedure CT, followed by cardiac catheterization. Overall, 23 morphologic parameters were manually extracted from cardiac CT images for each patient. The Mann-Whitney U or chi-square test was applied to select the most significant predictors. Six machine learning algorithms including logistic regression, Naive Bayes, random forest (RF), linear discriminant analysis, support vector machine, and K-nearest neighbor were used for modeling. These algorithms were independently trained on 100 train-validation random splits with a 3:1 ratio. Their average performance was evaluated by area under the curve (AUC), accuracy, sensitivity, and specificity.

RESULTS

Seven CT morphologic parameters were selected for modeling. RF obtained the best performance, with mean AUC of 0.840 (confidence interval [CI] 0.832-0.850) and 0.787 (95% CI 0.780-0.794); sensitivity of 0.815 (95% CI 0.797-0.833) and 0.778 (95% CI 0.767-0.788), specificity of 0.766 (95% CI 0.748-0.785) and 0.746 (95% CI 0.735-0.757); and accuracy of 0.782 (95% CI 0.771-0.793) and 0.756 (95% CI 0.748-0.764) in the training and validation cohorts, respectively.

CONCLUSIONS

The CT-based RF model demonstrates a good performance in the prediction of mPAP, which may reduce the need for right heart catheterization in post-Glenn shunt patients with suspected mPAP > 15 mmHg.

KEY POINTS

• Twenty-three candidate descriptors were manually extracted from cardiac computed tomography images, and seven of them were selected for subsequent modeling. • The random forest model presents the best predictive performance for pulmonary pressure among all methods. • The computed tomography-based machine learning model could predict post-Glenn shunt pulmonary pressure non-invasively.