Artificial intelligence in interventional radiology: state of the art

Artificial intelligence (AI) has demonstrated great potential in a wide variety of applications in interventional radiology (IR). Support for decision-making and outcome prediction, new functions and improvements in fluoroscopy, ultrasound, computed tomography, and magnetic resonance imaging, specifically in the field of IR, have all been investigated. Furthermore, AI represents a significant boost for fusion imaging and simulated reality, robotics, touchless software interactions, and virtual biopsy. The procedural nature, heterogeneity, and lack of standardisation slow down the process of adoption of AI in IR. Research in AI is in its early stages as current literature is based on pilot or proof of concept studies. The full range of possibilities is yet to be explored.Relevance statement Exploring AI's transformative potential, this article assesses its current applications and challenges in IR, offering insights into decision support and outcome prediction, imaging enhancements, robotics, and touchless interactions, shaping the future of patient care.Key points• AI adoption in IR is more complex compared to diagnostic radiology.• Current literature about AI in IR is in its early stages.• AI has the potential to revolutionise every aspect of IR.

Early prediction of ventricular peritoneal shunt dependency in aneurysmal subarachnoid haemorrhage patients by recurrent neural network-based machine learning using routine intensive care unit data

Aneurysmal subarachnoid haemorrhage (aSAH) can lead to complications such as acute hydrocephalic congestion. Treatment of this acute condition often includes establishing an external ventricular drainage (EVD). However, chronic hydrocephalus develops in some patients, who then require placement of a permanent ventriculoperitoneal (VP) shunt. The aim of this study was to employ recurrent neural network (RNN)-based machine learning techniques to identify patients who require VP shunt placement at an early stage. This retrospective single-centre study included all patients who were diagnosed with aSAH and treated in the intensive care unit (ICU) between November 2010 and May 2020 (n = 602). More than 120 parameters were analysed, including routine neurocritical care data, vital signs and blood gas analyses. Various machine learning techniques, including RNNs and gradient boosting machines, were evaluated for their ability to predict VP shunt dependency. VP-shunt dependency could be predicted using an RNN after just one day of ICU stay, with an AUC-ROC of 0.77 (CI: 0.75-0.79). The accuracy of the prediction improved after four days of observation (Day 4: AUC-ROC 0.81, CI: 0.79-0.84). At that point, the accuracy of the prediction was 76% (CI: 75.98-83.09%), with a sensitivity of 85% (CI: 83-88%) and a specificity of 74% (CI: 71-78%). RNN-based machine learning has the potential to predict VP shunt dependency on Day 4 after ictus in aSAH patients using routine data collected in the ICU. The use of machine learning may allow early identification of patients with specific therapeutic needs and accelerate the execution of required procedures.

Smart forecasting of artifacts in contrast-enhanced breast MRI before contrast agent administration

OBJECTIVES

To evaluate whether artifacts on contrast-enhanced (CE) breast MRI maximum intensity projections (MIPs) might already be forecast before gadolinium-based contrast agent (GBCA) administration during an ongoing examination by analyzing the unenhanced T1-weighted images acquired before the GBCA injection.

MATERIALS AND METHODS

This IRB-approved retrospective analysis consisted of n = 2884 breast CE MRI examinations after intravenous administration of GBCA, acquired with n = 4 different MRI devices at different field strengths (1.5 T/3 T) during clinical routine. CE-derived subtraction MIPs were used to conduct a multi-class multi-reader evaluation of the presence and severity of artifacts with three independent readers. An ensemble classifier (EC) of five DenseNet models was used to predict artifacts for the post-contrast subtraction MIPs, giving as the input source only the pre-contrast T1-weighted sequence. Thus, the acquisition directly preceded the GBCA injection. The area under ROC (AuROC) and diagnostics accuracy scores were used to assess the performance of the neural network in an independent holdout test set (n = 285).

RESULTS

After majority voting, potentially significant artifacts were detected in 53.6% (n = 1521) of all breast MRI examinations (age 49.6 ± 12.6 years). In the holdout test set (mean age 49.7 ± 11.8 years), at a specificity level of 89%, the EC could forecast around one-third of artifacts (sensitivity 31%) before GBCA administration, with an AuROC = 0.66.

CONCLUSION

This study demonstrates the capability of a neural network to forecast the occurrence of artifacts on CE subtraction data before the GBCA administration. If confirmed in larger studies, this might enable a workflow-blended approach to prevent breast MRI artifacts by implementing in-scan personalized predictive algorithms.

CLINICAL RELEVANCE STATEMENT

Some artifacts in contrast-enhanced breast MRI maximum intensity projections might be predictable before gadolinium-based contrast agent injection using a neural network.

KEY POINTS

• Potentially significant artifacts can be observed in a relevant proportion of breast MRI subtraction sequences after gadolinium-based contrast agent administration (GBCA). • Forecasting the occurrence of such artifacts in subtraction maximum intensity projections before GBCA administration for individual patients was feasible at 89% specificity, which allowed correctly predicting one in three future artifacts. • Further research is necessary to investigate the clinical value of such smart personalized imaging approaches.

Voxel-based morphometry in single subjects without a scanner-specific normal database using a convolutional neural network

OBJECTIVES

Reliable detection of disease-specific atrophy in individual T1w-MRI by voxel-based morphometry (VBM) requires scanner-specific normal databases (NDB), which often are not available. The aim of this retrospective study was to design, train, and test a deep convolutional neural network (CNN) for single-subject VBM without the need for a NDB (CNN-VBM).

MATERIALS AND METHODS

The training dataset comprised 8945 T1w scans from 65 different scanners. The gold standard VBM maps were obtained by conventional VBM with a scanner-specific NDB for each of the 65 scanners. CNN-VBM was tested in an independent dataset comprising healthy controls (n = 37) and subjects with Alzheimer's disease (AD, n = 51) or frontotemporal lobar degeneration (FTLD, n = 30). A scanner-specific NDB for the generation of the gold standard VBM maps was available also for the test set. The technical performance of CNN-VBM was characterized by the Dice coefficient of CNN-VBM maps relative to VBM maps from scanner-specific VBM. For clinical testing, VBM maps were categorized visually according to the clinical diagnoses in the test set by two independent readers, separately for both VBM methods.

RESULTS

The VBM maps from CNN-VBM were similar to the scanner-specific VBM maps (median Dice coefficient 0.85, interquartile range [0.81, 0.90]). Overall accuracy of the visual categorization of the VBM maps for the detection of AD or FTLD was 89.8% for CNN-VBM and 89.0% for scanner-specific VBM.

CONCLUSION

CNN-VBM without NDB provides a similar performance in the detection of AD- and FTLD-specific atrophy as conventional VBM.

CLINICAL RELEVANCE STATEMENT

A deep convolutional neural network for voxel-based morphometry eliminates the need of scanner-specific normal databases without relevant performance loss and, therefore, could pave the way for the widespread clinical use of voxel-based morphometry to support the diagnosis of neurodegenerative diseases.

KEY POINTS

• The need of normal databases is a barrier for widespread use of voxel-based brain morphometry. • A convolutional neural network achieved a similar performance for detection of atrophy than conventional voxel-based morphometry. • Convolutional neural networks can pave the way for widespread clinical use of voxel-based morphometry.

Deep learning performance for detection and classification of microcalcifications on mammography

BACKGROUND

Breast cancer screening through mammography is crucial for early detection, yet the demand for mammography services surpasses the capacity of radiologists. Artificial intelligence (AI) can assist in evaluating microcalcifications on mammography. We developed and tested an AI model for localizing and characterizing microcalcifications.

METHODS

Three expert radiologists annotated a dataset of mammograms using histology-based ground truth. The dataset was partitioned for training, validation, and testing. Three neural networks (AlexNet, ResNet18, and ResNet34) were trained and evaluated using specific metrics including receiver operating characteristics area under the curve (AUC), sensitivity, and specificity. The reported metrics were computed on the test set (10% of the whole dataset).

RESULTS

The dataset included 1,000 patients aged 21-73 years and 1,986 mammograms (180 density A, 220 density B, 380 density C, and 220 density D), with 389 malignant and 611 benign groups of microcalcifications. AlexNet achieved the best performance with 0.98 sensitivity, 0.89 specificity of, and 0.98 AUC for microcalcifications detection and 0.85 sensitivity, 0.89 specificity, and 0.94 AUC of for microcalcifications classification. For microcalcifications detection, ResNet18 and ResNet34 achieved 0.96 and 0.97 sensitivity, 0.91 and 0.90 specificity and 0.98 and 0.98 AUC, retrospectively. For microcalcifications classification, ResNet18 and ResNet34 exhibited 0.75 and 0.84 sensitivity, 0.85 and 0.84 specificity, and 0.88 and 0.92 AUC, respectively.

CONCLUSIONS

The developed AI models accurately detect and characterize microcalcifications on mammography.

RELEVANCE STATEMENT

AI-based systems have the potential to assist radiologists in interpreting microcalcifications on mammograms. The study highlights the importance of developing reliable deep learning models possibly applied to breast cancer screening.

KEY POINTS

• A novel AI tool was developed and tested to aid radiologists in the interpretation of mammography by accurately detecting and characterizing microcalcifications. • Three neural networks (AlexNet, ResNet18, and ResNet34) were trained, validated, and tested using an annotated dataset of 1,000 patients and 1,986 mammograms. • The AI tool demonstrated high accuracy in detecting/localizing and characterizing microcalcifications on mammography, highlighting the potential of AI-based systems to assist radiologists in the interpretation of mammograms.

U-net convolutional neural network applied to progressive fibrotic interstitial lung disease: Is progression at CT scan associated with a clinical outcome?

BACKGROUND

Computational advances in artificial intelligence have led to the recent emergence of U-Net convolutional neural networks (CNNs) applied to medical imaging. Our objectives were to assess the progression of fibrotic interstitial lung disease (ILD) using routine CT scans processed by a U-Net CNN developed by our research team, and to identify a progression threshold indicative of poor prognosis.

METHODS

CT scans and clinical history of 32 patients with idiopathic fibrotic ILDs were retrospectively reviewed. Successive CT scans were processed by the U-Net CNN and ILD quantification was obtained. Correlation between ILD and FVC changes was assessed. ROC curve was used to define a threshold of ILD progression rate (PR) to predict poor prognostic (mortality or lung transplantation). The PR threshold was used to compare the cohort survival with Kaplan Mayer curves and log-rank test.

RESULTS

The follow-up was 3.8 ± 1.5 years encompassing 105 CT scans, with 3.3 ± 1.1 CT scans per patient. A significant correlation between ILD and FVC changes was obtained (p = 0.004, ρ = -0.30 [95% CI: -0.16 to -0.45]). Sixteen patients (50%) experienced unfavorable outcome including 13 deaths and 3 lung transplantations. ROC curve analysis showed an aera under curve of 0.83 (p < 0.001), with an optimal cut-off PR value of 4%/year. Patients exhibiting a PR ≥ 4%/year during the first two years had a poorer prognosis (p = 0.001).

CONCLUSIONS

Applying a U-Net CNN to routine CT scan allowed identifying patients with a rapid progression and unfavorable outcome.

Deep learning-aided extraction of outer aortic surface from CT angiography scans of patients with Stanford type B aortic dissection

BACKGROUND

Guidelines recommend that aortic dimension measurements in aortic dissection should include the aortic wall. This study aimed to evaluate two-dimensional (2D)- and three-dimensional (3D)-based deep learning approaches for extraction of outer aortic surface in computed tomography angiography (CTA) scans of Stanford type B aortic dissection (TBAD) patients and assess the speed of different whole aorta (WA) segmentation approaches.

METHODS

A total of 240 patients diagnosed with TBAD between January 2007 and December 2019 were retrospectively reviewed for this study; 206 CTA scans from 206 patients with acute, subacute, or chronic TBAD acquired with various scanners in multiple different hospital units were included. Ground truth (GT) WAs for 80 scans were segmented by a radiologist using an open-source software. The remaining 126 GT WAs were generated via semi-automatic segmentation process in which an ensemble of 3D convolutional neural networks (CNNs) aided the radiologist. Using 136 scans for training, 30 for validation, and 40 for testing, 2D and 3D CNNs were trained to automatically segment WA. Main evaluation metrics for outer surface extraction and segmentation accuracy were normalized surface Dice (NSD) and Dice coefficient score (DCS), respectively.

RESULTS

2D CNN outperformed 3D CNN in NSD score (0.92 versus 0.90, p = 0.009), and both CNNs had equal DCS (0.96 versus 0.96, p = 0.110). Manual and semi-automatic segmentation times of one CTA scan were approximately 1 and 0.5 h, respectively.

CONCLUSIONS

Both CNNs segmented WA with high DCS, but based on NSD, better accuracy may be required before clinical application. CNN-based semi-automatic segmentation methods can expedite the generation of GTs.

RELEVANCE STATEMENT

Deep learning can speeds up the creation of ground truth segmentations. CNNs can extract the outer aortic surface in patients with type B aortic dissection.

KEY POINTS

• 2D and 3D convolutional neural networks (CNNs) can extract the outer aortic surface accurately. • Equal Dice coefficient score (0.96) was reached with 2D and 3D CNNs. • Deep learning can expedite the creation of ground truth segmentations.

Leveraging open dataset and transfer learning for accurate recognition of chronic pulmonary embolism from CT angiogram maximum intensity projection images

BACKGROUND

Early diagnosis of the potentially fatal but curable chronic pulmonary embolism (CPE) is challenging. We have developed and investigated a novel convolutional neural network (CNN) model to recognise CPE from CT pulmonary angiograms (CTPA) based on the general vascular morphology in two-dimensional (2D) maximum intensity projection images.

METHODS

A CNN model was trained on a curated subset of a public pulmonary embolism CT dataset (RSPECT) with 755 CTPA studies, including patient-level labels of CPE, acute pulmonary embolism (APE), or no pulmonary embolism. CPE patients with right-to-left-ventricular ratio (RV/LV) < 1 and APE patients with RV/LV ≥ 1 were excluded from the training. Additional CNN model selection and testing were done on local data with 78 patients without the RV/LV-based exclusion. We calculated area under the receiver operating characteristic curves (AUC) and balanced accuracies to evaluate the CNN performance.

RESULTS

We achieved a very high CPE versus no-CPE classification AUC 0.94 and balanced accuracy 0.89 on the local dataset using an ensemble model and considering CPE to be present in either one or both lungs.

CONCLUSIONS

We propose a novel CNN model with excellent predictive accuracy to differentiate chronic pulmonary embolism with RV/LV ≥ 1 from acute pulmonary embolism and non-embolic cases from 2D maximum intensity projection reconstructions of CTPA.

RELEVANCE STATEMENT

A DL CNN model identifies chronic pulmonary embolism from CTA with an excellent predictive accuracy.

KEY POINTS

• Automatic recognition of CPE from computed tomography pulmonary angiography was developed. • Deep learning was applied on two-dimensional maximum intensity projection images. • A large public dataset was used for training the deep learning model. • The proposed model showed an excellent predictive accuracy.

Benefits of automated gross tumor volume segmentation in head and neck cancer using multi-modality information

PURPOSE

Gross tumor volume (GTV) delineation for head and neck cancer (HNC) radiation therapy planning is time consuming and prone to interobserver variability (IOV). The aim of this study was (1) to develop an automated GTV delineation approach of primary tumor (GTVp) and pathologic lymph nodes (GTVn) based on a 3D convolutional neural network (CNN) exploiting multi-modality imaging input as required in clinical practice, and (2) to validate its accuracy, efficiency and IOV compared to manual delineation in a clinical setting.

METHODS

Two datasets were retrospectively collected from 150 clinical cases. CNNs were trained for GTV delineation with consensus delineation as ground truth, with either single (CT) or co-registered multi-modal (CT + PET or CT + MRI) imaging data as input. For validation, GTVs were delineated on 20 new cases by two observers, once manually, once by correcting the delineations generated by the CNN.

RESULTS

Both multi-modality CNNs performed better than the single-modality CNN and were selected for clinical validation. Mean Dice Similarity Coefficient (DSC) for (GTVp, GTVn) respectively between automated and manual delineations was (69%, 79%) for CT + PET and (59%,71%) for CT + MRI. Mean DSC between automated and corrected delineations was (81%,89%) for CT + PET and (69%,77%) for CT + MRI. Mean DSC between observers was (76%,86%) for manual delineations and (95%,96%) for corrected delineations, indicating a significant decrease in IOV (p < 10^-5), while efficiency increased significantly (48%, p < 10^-5).

CONCLUSION

Multi-modality automated delineation of GTV of HNC was shown to be more efficient and consistent compared to manual delineation in a clinical setting and beneficial over a single-modality approach.

Performance of a deep convolutional neural network for MRI-based vertebral body measurements and insufficiency fracture detection

OBJECTIVES

The aim is to validate the performance of a deep convolutional neural network (DCNN) for vertebral body measurements and insufficiency fracture detection on lumbar spine MRI.

METHODS

This retrospective analysis included 1000 vertebral bodies in 200 patients (age 75.2 ± 9.8 years) who underwent lumbar spine MRI at multiple institutions. 160/200 patients had ≥ one vertebral body insufficiency fracture, 40/200 had no fracture. The performance of the DCNN and that of two fellowship-trained musculoskeletal radiologists in vertebral body measurements (anterior/posterior height, extent of endplate concavity, vertebral angle) and evaluation for insufficiency fractures were compared. Statistics included (a) interobserver reliability metrics using intraclass correlation coefficient (ICC), kappa statistics, and Bland-Altman analysis, and (b) diagnostic performance metrics (sensitivity, specificity, accuracy). A statistically significant difference was accepted if the 95% confidence intervals did not overlap.

RESULTS

The inter-reader agreement between radiologists and the DCNN was excellent for vertebral body measurements, with ICC values of > 0.94 for anterior and posterior vertebral height and vertebral angle, and good to excellent for superior and inferior endplate concavity with ICC values of 0.79-0.85. The performance of the DCNN in fracture detection yielded a sensitivity of 0.941 (0.903-0.968), specificity of 0.969 (0.954-0.980), and accuracy of 0.962 (0.948-0.973). The diagnostic performance of the DCNN was independent of the radiological institution (accuracy 0.964 vs. 0.960), type of MRI scanner (accuracy 0.957 vs. 0.964), and magnetic field strength (accuracy 0.966 vs. 0.957).

CONCLUSIONS

A DCNN can achieve high diagnostic performance in vertebral body measurements and insufficiency fracture detection on heterogeneous lumbar spine MRI.

KEY POINTS

• A DCNN has the potential for high diagnostic performance in measuring vertebral bodies and detecting insufficiency fractures of the lumbar spine.

Generation of synthetic ground glass nodules using generative adversarial networks (GANs)

BACKGROUND

Data shortage is a common challenge in developing computer-aided diagnosis systems. We developed a generative adversarial network (GAN) model to generate synthetic lung lesions mimicking ground glass nodules (GGNs).

METHODS

We used 216 computed tomography images with 340 GGNs from the Lung Image Database Consortium and Image Database Resource Initiative database. A GAN model retrieving information from the whole image and the GGN region was built. The generated samples were evaluated with visual Turing test performed by four experienced radiologists or pulmonologists. Radiomic features were compared between real and synthetic nodules. Performances were evaluated by area under the curve (AUC) at receiver operating characteristic analysis. In addition, we trained a classification model (ResNet) to investigate whether the synthetic GGNs can improve the performances algorithm and how performances changed as a function of labelled data used in training.

RESULTS

Of 51 synthetic GGNs, 19 (37%) were classified as real by clinicians. Of 93 radiomic features, 58 (62.4%) showed no significant difference between synthetic and real GGNs (p ≥ 0.052). The discrimination performances of physicians (AUC 0.68) and radiomics (AUC 0.66) were similar, with no-significantly different (p = 0.23), but clinicians achieved a better accuracy (AUC 0.74) than radiomics (AUC 0.62) (p < 0.001). The classification model trained on datasets with synthetic data performed better than models without the addition of synthetic data.

CONCLUSIONS

GAN has promising potential for generating GGNs. Through similar AUC, clinicians achieved better ability to diagnose whether the data is synthetic than radiomics.

Diagnosis of in vivo vertical root fracture using deep learning on cone-beam CT images

Objectives

Evaluating the diagnostic efficiency of deep learning models to diagnose vertical root fracture in vivo on cone-beam CT (CBCT) images.

Materials and methods

The CBCT images of 276 teeth (138 VRF teeth and 138 non-VRF teeth) were enrolled and analyzed retrospectively. The diagnostic results of these teeth were confirmed by two chief radiologists. There were two experimental groups: auto-selection group and manual selection group. A total of 552 regions of interest of teeth were cropped in manual selection group and 1118 regions of interest of teeth were cropped in auto-selection group. Three deep learning networks (ResNet50, VGG19 and DenseNet169) were used for diagnosis (3:1 for training and testing). The diagnostic efficiencies (accuracy, sensitivity, specificity, and area under the curve (AUC)) of three networks were calculated in two experiment groups. Meanwhile, 552 teeth images in manual selection group were diagnosed by a radiologist. The diagnostic efficiencies of the three deep learning network models in two experiment groups and the radiologist were calculated.

Results

In manual selection group, ResNet50 presented highest accuracy and sensitivity for diagnosing VRF teeth. The accuracy, sensitivity, specificity and AUC was 97.8%, 97.0%, 98.5%, and 0.99, the radiologist presented accuracy, sensitivity, and specificity as 95.3%, 96.4 and 94.2%. In auto-selection group, ResNet50 presented highest accuracy and sensitivity for diagnosing VRF teeth, the accuracy, sensitivity, specificity and AUC was 91.4%, 92.1%, 90.7% and 0.96.

Conclusion

In manual selection group, ResNet50 presented higher diagnostic efficiency in diagnosis of in vivo VRF teeth than VGG19, DensenNet169 and radiologist with 2 years of experience. In auto-selection group, Resnet50 also presented higher diagnostic efficiency in diagnosis of in vivo VRF teeth than VGG19 and DensenNet169. This makes it a promising auxiliary diagnostic technique to screen for VRF teeth.

Performance of a 3D convolutional neural network in the detection of hypoperfusion at CT pulmonary angiography in patients with chronic pulmonary embolism: a feasibility study

Background

Chronic pulmonary embolism (CPE) is a life-threatening disease easily misdiagnosed on computed tomography. We investigated a three-dimensional convolutional neural network (CNN) algorithm for detecting hypoperfusion in CPE from computed tomography pulmonary angiography (CTPA).

Methods

Preoperative CTPA of 25 patients with CPE and 25 without pulmonary embolism were selected. We applied a 48%–12%–40% training-validation-testing split (12 positive and 12 negative CTPA volumes for training, 3 positives and 3 negatives for validation, 10 positives and 10 negatives for testing). The median number of axial images per CTPA was 335 (min–max, 111–570). Expert manual segmentations were used as training and testing targets. The CNN output was compared to a method in which a Hounsfield unit (HU) threshold was used to detect hypoperfusion. Receiver operating characteristic area under the curve (AUC) and Matthew correlation coefficient (MCC) were calculated with their 95% confidence interval (CI).

Results

The predicted segmentations of CNN showed AUC 0.87 (95% CI 0.82–0.91), those of HU-threshold method 0.79 (95% CI 0.74–0.84). The optimal global threshold values were CNN output probability ≥ 0.37 and ≤ -850 HU. Using these values, MCC was 0.46 (95% CI 0.29–0.59) for CNN and 0.35 (95% CI 0.18–0.48) for HU-threshold method (average difference in MCC in the bootstrap samples 0.11 (95% CI 0.05–0.16). A high CNN prediction probability was a strong predictor of CPE.

Conclusions

We proposed a deep learning method for detecting hypoperfusion in CPE from CTPA. This model may help evaluating disease extent and supporting treatment planning.

Supporting the classification of patients in public hospitals in Chile by designing, deploying and validating a system based on natural language processing

Background

In Chile, a patient needing a specialty consultation or surgery has to first be referred by a general practitioner, then placed on a waiting list. The Explicit Health Guarantees (GES in Spanish) ensures, by law, the maximum time to solve 85 health problems. Usually, a health professional manually verifies if each referral, written in natural language, corresponds or not to a GES-covered disease. An error in this classification is catastrophic for patients, as it puts them on a non-prioritized waiting list, characterized by prolonged waiting times.

Methods

To support the manual process, we developed and deployed a system that automatically classifies referrals as GES-covered or not using historical data. Our system is based on word embeddings specially trained for clinical text produced in Chile. We used a vector representation of the reason for referral and patient's age as features for training machine learning models using human-labeled historical data. We constructed a ground truth dataset combining classifications made by three healthcare experts, which was used to validate our results.

Results

The best performing model over ground truth reached an AUC score of 0.94, with a weighted F1-score of 0.85 (0.87 in precision and 0.86 in recall). During seven months of continuous and voluntary use, the system has amended 87 patient misclassifications.

Conclusion

This system is a result of a collaboration between technical and clinical experts, and the design of the classifier was custom-tailored for a hospital's clinical workflow, which encouraged the voluntary use of the platform. Our solution can be easily expanded across other hospitals since the registry is uniform in Chile.

Artificial intelligence-aided CT segmentation for body composition analysis: a validation study

Background

Body composition is associated with survival outcome in oncological patients, but it is not routinely calculated. Manual segmentation of subcutaneous adipose tissue (SAT) and muscle is time-consuming and therefore limited to a single CT slice. Our goal was to develop an artificial-intelligence (AI)-based method for automated quantification of three-dimensional SAT and muscle volumes from CT images.

Methods

Ethical approvals from Gothenburg and Lund Universities were obtained. Convolutional neural networks were trained to segment SAT and muscle using manual segmentations on CT images from a training group of 50 patients. The method was applied to a separate test group of 74 cancer patients, who had two CT studies each with a median interval between the studies of 3 days. Manual segmentations in a single CT slice were used for comparison. The accuracy was measured as overlap between the automated and manual segmentations.

Results

The accuracy of the AI method was 0.96 for SAT and 0.94 for muscle. The average differences in volumes were significantly lower than the corresponding differences in areas in a single CT slice: 1.8% versus 5.0% (p < 0.001) for SAT and 1.9% versus 3.9% (p < 0.001) for muscle. The 95% confidence intervals for predicted volumes in an individual subject from the corresponding single CT slice areas were in the order of ± 20%.

Conclusions

The AI-based tool for quantification of SAT and muscle volumes showed high accuracy and reproducibility and provided a body composition analysis that is more relevant than manual analysis of a single CT slice.

Machine learning applied on chest x-ray can aid in the diagnosis of COVID-19: a first experience from Lombardy, Italy

BACKGROUND

We aimed to train and test a deep learning classifier to support the diagnosis of coronavirus disease 2019 (COVID-19) using chest x-ray (CXR) on a cohort of subjects from two hospitals in Lombardy, Italy.

METHODS

We used for training and validation an ensemble of ten convolutional neural networks (CNNs) with mainly bedside CXRs of 250 COVID-19 and 250 non-COVID-19 subjects from two hospitals (Centres 1 and 2). We then tested such system on bedside CXRs of an independent group of 110 patients (74 COVID-19, 36 non-COVID-19) from one of the two hospitals. A retrospective reading was performed by two radiologists in the absence of any clinical information, with the aim to differentiate COVID-19 from non-COVID-19 patients. Real-time polymerase chain reaction served as the reference standard.

RESULTS

At 10-fold cross-validation, our deep learning model classified COVID-19 and non-COVID-19 patients with 0.78 sensitivity (95% confidence interval [CI] 0.74-0.81), 0.82 specificity (95% CI 0.78-0.85), and 0.89 area under the curve (AUC) (95% CI 0.86-0.91). For the independent dataset, deep learning showed 0.80 sensitivity (95% CI 0.72-0.86) (59/74), 0.81 specificity (29/36) (95% CI 0.73-0.87), and 0.81 AUC (95% CI 0.73-0.87). Radiologists' reading obtained 0.63 sensitivity (95% CI 0.52-0.74) and 0.78 specificity (95% CI 0.61-0.90) in Centre 1 and 0.64 sensitivity (95% CI 0.52-0.74) and 0.86 specificity (95% CI 0.71-0.95) in Centre 2.

CONCLUSIONS

This preliminary experience based on ten CNNs trained on a limited training dataset shows an interesting potential of deep learning for COVID-19 diagnosis. Such tool is in training with new CXRs to further increase its performance.

Machine learning applied on chest x-ray can aid in the diagnosis of COVID-19: a first experience from Lombardy, Italy

BACKGROUND

We aimed to train and test a deep learning classifier to support the diagnosis of coronavirus disease 2019 (COVID-19) using chest x-ray (CXR) on a cohort of subjects from two hospitals in Lombardy, Italy.

METHODS

We used for training and validation an ensemble of ten convolutional neural networks (CNNs) with mainly bedside CXRs of 250 COVID-19 and 250 non-COVID-19 subjects from two hospitals (Centres 1 and 2). We then tested such system on bedside CXRs of an independent group of 110 patients (74 COVID-19, 36 non-COVID-19) from one of the two hospitals. A retrospective reading was performed by two radiologists in the absence of any clinical information, with the aim to differentiate COVID-19 from non-COVID-19 patients. Real-time polymerase chain reaction served as the reference standard.

RESULTS

At 10-fold cross-validation, our deep learning model classified COVID-19 and non-COVID-19 patients with 0.78 sensitivity (95% confidence interval [CI] 0.74-0.81), 0.82 specificity (95% CI 0.78-0.85), and 0.89 area under the curve (AUC) (95% CI 0.86-0.91). For the independent dataset, deep learning showed 0.80 sensitivity (95% CI 0.72-0.86) (59/74), 0.81 specificity (29/36) (95% CI 0.73-0.87), and 0.81 AUC (95% CI 0.73-0.87). Radiologists' reading obtained 0.63 sensitivity (95% CI 0.52-0.74) and 0.78 specificity (95% CI 0.61-0.90) in Centre 1 and 0.64 sensitivity (95% CI 0.52-0.74) and 0.86 specificity (95% CI 0.71-0.95) in Centre 2.

CONCLUSIONS

This preliminary experience based on ten CNNs trained on a limited training dataset shows an interesting potential of deep learning for COVID-19 diagnosis. Such tool is in training with new CXRs to further increase its performance.

Natural history of pathologically confirmed pulmonary subsolid nodules with deep learning-assisted nodule segmentation

OBJECTIVE

To explore the natural history of pulmonary subsolid nodules (SSNs) with different pathological types by deep learning-assisted nodule segmentation.

METHODS

Between June 2012 and June 2019, 95 resected SSNs with preoperative long-term follow-up were enrolled in this retrospective study. SSN detection and segmentation were performed on preoperative follow-up CTs using the deep learning-based Dr. Wise system. SSNs were categorized into invasive adenocarcinoma (IAC, n = 47) and non-IAC (n = 48) groups; according to the interval change during the preoperative follow-up, SSNs were divided into growth (n = 68), nongrowth (n = 22), and new emergence (n = 5) groups. We analyzed the cumulative percentages and pattern of SSN growth and identified significant factors for IAC diagnosis and SSN growth.

RESULTS

The mean preoperative follow-up was 42.1 ± 17.0 months. More SSNs showed growth or new emergence in the IAC than in the non-IAC group (89.4% vs. 64.6%, p = 0.009). Volume doubling time was non-significantly shorter for IACs than for non-IACs (1436.0 ± 1188.2 vs. 2087.5 ± 1799.7 days, p = 0.077). Median mass doubling time was significantly shorter for IACs than for non-IACs (821.7 vs. 1944.1 days, p = 0.001). Lobulated sign (p = 0.002) and SSN mass (p = 0.004) were significant factors for differentiating IACs. IACs showed significantly higher cumulative growth percentages than non-IACs in the first 70 months of follow-up. The growth pattern of SSNs may conform to the exponential model. The initial volume (p = 0.042) was a predictor for SSN growth.

CONCLUSIONS

IACs appearing as SSNs showed an indolent course. The mean growth rate was larger for IACs than for non-IACs. SSNs with larger initial volume are more likely to grow.

KEY POINTS

• Invasive adenocarcinomas (IACs) appearing as subsolid nodules (SSNs), with a mean volume doubling time (VDT) of 1436.0 ± 1188.2 days and median mass doubling time (MDT) of 821.7 days, showed an indolent course. • The VDT was shorter for IACs than for non-IACs (1436.0 ± 1188.2 vs. 2087.5 ± 1799.7 days), but the difference was not significant (p = 0.077). The median MDT was significantly shorter for IACs than for non-IACs (821.7 vs. 1944.1 days, p = 0.001). • SSNs with lobulated sign and larger mass (> 390.5 mg) may very likely be IACs. SSNs with larger initial volume are more likely to grow.

Deep learning for elective neck delineation: More consistent and time efficient

BACKGROUND/PURPOSE

Delineation of the lymph node levels of the neck for irradiation of the elective clinical target volume in head and neck cancer (HNC) patients is time consuming and prone to interobserver variability (IOV), although international consensus guidelines exist. The aim of this study was to develop and validate a 3D convolutional neural network (CNN) for semi-automated delineation of all nodal neck levels, focussing on delineation accuracy, efficiency and consistency compared to manual delineation.

MATERIAL/METHODS

The CNN was trained on a clinical dataset of 69 HNC patients. For validation, 17 lymph node levels were manually delineated in 16 new patients by two observers, independently, using international consensus guidelines. Automated delineations were generated by applying the CNN and were subsequently corrected by both observers separately as needed for clinical acceptance. Both delineations were performed two weeks apart and blinded to each other. IOV was quantified using Dice similarity coefficient (DSC), mean surface distance (MSD) and Hausdorff distance (HD). To assess automated delineation accuracy, agreement between automated and corrected delineations were evaluated using the same measures. To assess efficiency, the time taken for manual and corrected delineations were compared. In a second step, only the clinically relevant neck levels were selected and delineated, once again manually and by applying and correcting the network.

RESULTS

When all lymph node levels were delineated, time taken for correcting automated delineations compared to manual delineations was significantly shorter for both observers (mean: 35 vs 52 min, p < 10^-5). Based on DSC, automated delineation agreed best with corrected delineation for lymph node levels Ib, II-IVa, VIa, VIb, VIIa, VIIb (DSC >85%). Manual corrections necessary for clinical acceptance were 1.4 mm MSD on average and were especially low (<1mm) for levels II-IVa, VIa, VIIa and VIIb. IOV was significantly smaller with automated compared to manual delineations (MSD: 1.4 mm vs 2.5 mm, p < 10^-11). When delineating only the clinically relevant neck levels, the correction time was also significantly shorter (mean: 8 vs 15 min, p < 10^-5). Based on DSC, automated delineation agreed very well with corrected delineation (DSC > 87%). Manual corrections necessary for clinical acceptance were 1.3 mm MSD on average. IOV was significantly smaller with automated compared to manual delineations (MSD: 0.8 mm vs 2.3 mm, p < 10^-3).

CONCLUSION

The CNN developed for automated delineation of the elective lymph node levels in the neck in HNC was shown to be more efficient and consistent compared to manual delineation, which justifies its implementation in clinical practice.

[Artificial intelligence and machine learning in oncologic imaging]

Machine learning (ML) is entering many areas of society, including medicine. This transformation has the potential to drastically change medicine and medical practice. These aspects become particularly clear when considering the different stages of oncologic patient care and the involved interdisciplinary and intermodality interactions. In recent publications, computers-in collaboration with humans or alone-have been outperforming humans regarding tumor identification, tumor classification, estimating prognoses, and evaluation of treatments. In addition, ML algorithms, e.g., artificial neural networks (ANNs), which constitute the drivers behind many of the latest achievements in ML, can deliver this level of performance in a reproducible, fast, and inexpensive manner. In the future, artificial intelligence applications will become an integral part of the medical profession and offer advantages for oncologic diagnostics and treatment.

Deep learning reconstruction of drip-infusion cholangiography acquired with ultra-high-resolution computed tomography

PURPOSE

Deep learning reconstruction (DLR) introduces deep convolutional neural networks into the reconstruction flow. We examined the clinical applicability of drip-infusion cholangiography (DIC) acquired on an ultra-high-resolution CT (U-HRCT) scanner reconstructed with DLR in comparison to hybrid and model-based iterative reconstruction (hybrid-IR, MBIR).

METHODS

This retrospective, single-institution study included 30 patients seen between January 2018 and November 2019. A radiologist recorded the standard deviation of attenuation in the paraspinal muscle as the image noise and calculated the contrast-to-noise ratio (CNR) in the common bile duct. The overall visual image quality of the bile duct on thick-slab maximum intensity projections was assessed by two other radiologists and graded on a 5-point confidence scale ranging from 1 (not delineated) to 5 (clearly delineated). The difference among hybrid-IR, MBIR, and DLR images was compared.

RESULTS

The image noise was significantly lower on DLR than hybrid-IR and MBIR images and the CNR and the overall visual image quality of the bile duct were significantly higher on DLR than on hybrid-IR and MBIR images (all: p < 0.001).

CONCLUSION

DLR resulted in significant quantitative and qualitative improvement of DIC acquired with U-HRCT.

Clinically significant prostate cancer detection and segmentation in low-risk patients using a convolutional neural network on multi-parametric MRI

Objectives

To develop an automatic method for identification and segmentation of clinically significant prostate cancer in low-risk patients and to evaluate the performance in a routine clinical setting.

Methods

A consecutive cohort (n = 292) from a prospective database of low-risk patients eligible for the active surveillance was selected. A 3-T multi-parametric MRI at 3 months after inclusion was performed. Histopathology from biopsies was used as reference standard. MRI positivity was defined as PI-RADS score ≥ 3, histopathology positivity was defined as ISUP grade ≥ 2. The selected cohort contained four patient groups: (1) MRI-positive targeted biopsy-positive (n = 116), (2) MRI-negative systematic biopsy-negative (n = 55), (3) MRI-positive targeted biopsy-negative (n = 113), (4) MRI-negative systematic biopsy-positive (n = 8). Group 1 was further divided into three sets and a 3D convolutional neural network was trained using different combinations of these sets. Two MRI sequences (T2w, b = 800 DWI) and the ADC map were used as separate input channels for the model. After training, the model was evaluated on the remaining group 1 patients together with the patients of groups 2 and 3 to identify and segment clinically significant prostate cancer.

Results

The average sensitivity achieved was 82–92% at an average specificity of 43–76% with an area under the curve (AUC) of 0.65 to 0.89 for different lesion volumes ranging from > 0.03 to > 0.5 cc.

Conclusions

The proposed deep learning computer-aided method yields promising results in identification and segmentation of clinically significant prostate cancer and in confirming low-risk cancer (ISUP grade ≤ 1) in patients on active surveillance.

Key Points

• Clinically significant prostate cancer identification and segmentation on multi-parametric MRI is feasible in low-risk patients using a deep neural network.

• The deep neural network for significant prostate cancer localization performs better for lesions with larger volumes sizes (> 0.5 cc) as compared to small lesions (> 0.03 cc).

• For the evaluation of automatic prostate cancer segmentation methods in the active surveillance cohort, the large discordance group (MRI positive, targeted biopsy negative) should be included.

Electronic supplementary material

The online version of this article (10.1007/s00330-020-07008-z) contains supplementary material, which is available to authorized users.

Multiphase CT-based prediction of Child-Pugh classification: a machine learning approach

Background

To evaluate whether machine learning algorithms allow the prediction of Child-Pugh classification on clinical multiphase computed tomography (CT).

Methods

A total of 259 patients who underwent diagnostic abdominal CT (unenhanced, contrast-enhanced arterial, and venous phases) were included in this retrospective study. Child-Pugh scores were determined based on laboratory and clinical parameters. Linear regression (LR), Random Forest (RF), and convolutional neural network (CNN) algorithms were used to predict the Child-Pugh class. Their performances were compared to the prediction of experienced radiologists (ERs). Spearman correlation coefficients and accuracy were assessed for all predictive models. Additionally, a binary classification in low disease severity (Child-Pugh class A) and advanced disease severity (Child-Pugh class ≥ B) was performed.

Results

Eleven imaging features exhibited a significant correlation when adjusted for multiple comparisons with Child-Pugh class. Significant correlations between predicted and measured Child-Pugh classes were observed (ρ_LA = 0.35, ρ_RF = 0.32, ρ_CNN = 0.51, ρ_ERs = 0.60; p < 0.001). Significantly better accuracies for the prediction of Child-Pugh classes versus no-information rate were found for CNN and ERs (p ≤ 0.034), not for LR and RF (p ≥ 0.384). For binary severity classification, the area under the curve at receiver operating characteristic analysis was significantly lower (p ≤ 0.042) for LR (0.71) and RF (0.69) than for CNN (0.80) and ERs (0.76), without significant differences between CNN and ERs (p = 0.144).

Conclusions

The performance of a CNN in assessing Child-Pugh class based on multiphase abdominal CT images is comparable to that of ERs.

Evaluation of acute pulmonary embolism and clot burden on CTPA with deep learning

OBJECTIVES

To take advantage of the deep learning algorithms to detect and calculate clot burden of acute pulmonary embolism (APE) on computed tomographic pulmonary angiography (CTPA).

MATERIALS AND METHODS

The training set in this retrospective study consisted of 590 patients (460 with APE and 130 without APE) who underwent CTPA. A fully deep learning convolutional neural network (DL-CNN), called U-Net, was trained for the segmentation of clot. Additionally, an in-house validation set consisted of 288 patients (186 with APE and 102 without APE). In this study, we set different probability thresholds to test the performance of U-Net for the clot detection and selected sensitivity, specificity, and area under the curve (AUC) as the metrics of performance evaluation. Furthermore, we investigated the relationship between the clot burden assessed by the Qanadli score, Mastora score, and other imaging parameters on CTPA and the clot burden calculated by the DL-CNN model.

RESULTS

There was no statistically significant difference in AUCs with the different probability thresholds. When the probability threshold for segmentation was 0.1, the sensitivity and specificity of U-Net in detecting clot respectively were 94.6% and 76.5% while the AUC was 0.926 (95% CI 0.884-0.968). Moreover, this study displayed that the clot burden measured with U-Net was significantly correlated with the Qanadli score (r = 0.819, p < 0.001), Mastora score (r = 0.874, p < 0.001), and right ventricular functional parameters on CTPA.

CONCLUSIONS

DL-CNN achieved a high AUC for the detection of pulmonary emboli and can be applied to quantitatively calculate the clot burden of APE patients, which may contribute to reducing the workloads of clinicians.

KEY POINTS

• Deep learning can detect APE with a good performance and efficiently calculate the clot burden to reduce the physicians' workload. • Clot burden measured with deep learning highly correlates with Qanadli and Mastora scores of CTPA. • Clot burden measured with deep learning correlates with parameters of right ventricular function on CTPA.

Automated volumetric assessment with artificial neural networks might enable a more accurate assessment of disease burden in patients with multiple sclerosis

OBJECTIVES

Patients with multiple sclerosis (MS) regularly undergo MRI for assessment of disease burden. However, interpretation may be time consuming and prone to intra- and interobserver variability. Here, we evaluate the potential of artificial neural networks (ANN) for automated volumetric assessment of MS disease burden and activity on MRI.

METHODS

A single-institutional dataset with 334 MS patients (334 MRI exams) was used to develop and train an ANN for automated identification and volumetric segmentation of T2/FLAIR-hyperintense and contrast-enhancing (CE) lesions. Independent testing was performed in a single-institutional longitudinal dataset with 82 patients (266 MRI exams). We evaluated lesion detection performance (F1 scores), lesion segmentation agreement (DICE coefficients), and lesion volume agreement (concordance correlation coefficients [CCC]). Independent evaluation was performed on the public ISBI-2015 challenge dataset.

RESULTS

The F1 score was maximized in the training set at a detection threshold of 7 mm³ for T2/FLAIR lesions and 14 mm³ for CE lesions. In the training set, mean F1 scores were 0.867 for T2/FLAIR lesions and 0.636 for CE lesions, as compared to 0.878 for T2/FLAIR lesions and 0.715 for CE lesions in the test set. Using these thresholds, the ANN yielded mean DICE coefficients of 0.834 and 0.878 for segmentation of T2/FLAIR and CE lesions in the training set (fivefold cross-validation). Corresponding DICE coefficients in the test set were 0.846 for T2/FLAIR lesions and 0.908 for CE lesions, and the CCC was ≥ 0.960 in each dataset.

CONCLUSIONS

Our results highlight the capability of ANN for quantitative state-of-the-art assessment of volumetric lesion load on MRI and potentially enable a more accurate assessment of disease burden in patients with MS.

KEY POINTS

• Artificial neural networks (ANN) can accurately detect and segment both T2/FLAIR and contrast-enhancing MS lesions in MRI data. • Performance of the ANN was consistent in a clinically derived dataset, with patients presenting all possible disease stages in MRI scans acquired from standard clinical routine rather than with high-quality research sequences. • Computer-aided evaluation of MS with ANN could streamline both clinical and research procedures in the volumetric assessment of MS disease burden as well as in lesion detection.

Deep convolutional neural network-based detection of meniscus tears: comparison with radiologists and surgery as standard of reference

OBJECTIVE

To clinically validate a fully automated deep convolutional neural network (DCNN) for detection of surgically proven meniscus tears.

MATERIALS AND METHODS

One hundred consecutive patients were retrospectively included, who underwent knee MRI and knee arthroscopy in our institution. All MRI were evaluated for medial and lateral meniscus tears by two musculoskeletal radiologists independently and by DCNN. Included patients were not part of the training set of the DCNN. Surgical reports served as the standard of reference. Statistics included sensitivity, specificity, accuracy, ROC curve analysis, and kappa statistics.

RESULTS

Fifty-seven percent (57/100) of patients had a tear of the medial and 24% (24/100) of the lateral meniscus, including 12% (12/100) with a tear of both menisci. For medial meniscus tear detection, sensitivity, specificity, and accuracy were for reader 1: 93%, 91%, and 92%, for reader 2: 96%, 86%, and 92%, and for the DCNN: 84%, 88%, and 86%. For lateral meniscus tear detection, sensitivity, specificity, and accuracy were for reader 1: 71%, 95%, and 89%, for reader 2: 67%, 99%, and 91%, and for the DCNN: 58%, 92%, and 84%. Sensitivity for medial meniscus tears was significantly different between reader 2 and the DCNN (p = 0.039), and no significant differences existed for all other comparisons (all p ≥ 0.092). The AUC-ROC of the DCNN was 0.882, 0.781, and 0.961 for detection of medial, lateral, and overall meniscus tear. Inter-reader agreement was very good for the medial (kappa = 0.876) and good for the lateral meniscus (kappa = 0.741).

CONCLUSION

DCNN-based meniscus tear detection can be performed in a fully automated manner with a similar specificity but a lower sensitivity in comparison with musculoskeletal radiologists.

Fully automated convolutional neural network-based affine algorithm improves liver registration and lesion co-localization on hepatobiliary phase T1-weighted MR images

BACKGROUND

Liver alignment between series/exams is challenged by dynamic morphology or variability in patient positioning or motion. Image registration can improve image interpretation and lesion co-localization. We assessed the performance of a convolutional neural network algorithm to register cross-sectional liver imaging series and compared its performance to manual image registration.

METHODS

Three hundred fourteen patients, including internal and external datasets, who underwent gadoxetate disodium-enhanced magnetic resonance imaging for clinical care from 2011 to 2018, were retrospectively selected. Automated registration was applied to all 2,663 within-patient series pairs derived from these datasets. Additionally, 100 within-patient series pairs from the internal dataset were independently manually registered by expert readers. Liver overlap, image correlation, and intra-observation distances for manual versus automated registrations were compared using paired t tests. Influence of patient demographics, imaging characteristics, and liver uptake function was evaluated using univariate and multivariate mixed models.

RESULTS

Compared to the manual, automated registration produced significantly lower intra-observation distance (p < 0.001) and higher liver overlap and image correlation (p < 0.001). Intra-exam automated registration achieved 0.88 mean liver overlap and 0.44 mean image correlation for the internal dataset and 0.91 and 0.41, respectively, for the external dataset. For inter-exam registration, mean overlap was 0.81 and image correlation 0.41. Older age, female sex, greater inter-series time interval, differing uptake, and greater voxel size differences independently reduced automated registration performance (p ≤ 0.020).

CONCLUSION

A fully automated algorithm accurately registered the liver within and between examinations, yielding better liver and focal observation co-localization compared to manual registration.

Computer algorithm can match physicians' decisions about blood transfusions

BACKGROUND

Checking appropriateness of blood transfusion for quality assurance required enormous usage of time and human resources from the healthcare system. We report here a new machine learning algorithm for checking blood transfusion quality.

MATERIALS AND METHODS

The multilayer perceptron neural network (MLPNN) was designed to learn an expert's judgement from 4946 clinical cases. The accuracy in predicting the blood transfusion was then reported.

RESULTS

We achieved a 96.8% overall accuracy rate, with a 99% match rate to the experts' judgement on those appropriate cases and 90.9% on the inappropriate cases.

CONCLUSIONS

Machine learning algorithm can accurately match to human judgement by feeding in pre-surgical information and key laboratory variables.

[Study on the method of microelectrodes implantation of artificial facial nerve prosthesis in closed mouth of orbicularis oris muscle in monkeys with facial nerve paralysis]

Objective: To explore the optimal method of microelectrode implantation that can produce efficient mouth closure with microelectrode for orbicularis oris muscle (OOM) in rhesus monkeys with unilateral peripheral facial paralysis (UPFP) in order to provide basis for the research and development of artificial facial nerve prosthesis (AFNP). Methods: Right lateral peripheral facial paralysis model on four healthy rhesus monkeys (two males and two femles, aged 5-6 years, weighed 2.0-3.0 kg) were prepared. AFNP electric stimulation was used to induce closed-mouth reaction of the affected OOM with a one-way rectangular pulse, 50 Hz frequency and 0.2 ms pulse width in vitro. Around the affected lateral OOM, four stimulus electrodes implantation positions were selected at the upper lip (position A), the lower lip (position B), the connection with the corner of the mouth to the ipsilateral tragus (position C), and the horizontal line of the mouth angle (position D). According to the different implantation positions of three stimulation electrodes on the stimulation side of AFNP and the results of our previous study, six groups of microelectrode implantation methods were designed. In Group A, two microelectrodes were implanted at position A and one microelectrode was implanted at position B; in Group B, one microelectrode was implanted at position A, B and C respectively; in Group C, one microelectrode was implanted at position A and two microelectrodes were implanted at position B; in Group D, one microelectrode was implanted at position A, B and D respectively; in Group E, one microelectrode was implanted at position A, C and D respectively; in Group F, one microelectrode was implanted at position B, C and D respectively. The minimum stimulating current (threshold current) required for effective mouth closure were recorded. The threshold and peak current values were compared using one-way ANOVA and LSD-t multiple comparisons. Results: The microelectrodes of the AFNP stimulating side in Group E and F failed to induce a smooth mouth closure. The microelectrodes in A, B, C and D group induced smooth mouth closure. The threshold current value of OOM contraction on affected side in the Group A, B, C, and D were (1.35±0.05), (1.02±0.04), (1.40±0.04) and (1.10±0.02) mA, respectively (F=295.302, P<0.001), with the lowest value in Group B and there was significant difference between the current value in Group B and those in the other groups (all P<0.05). The peak current value of OOM contraction on affected side in the four groups were (3.95±0.02), (2.95±0.03), (3.99±0.05) and (3.51±0.01) mA, respectively (F=1 014.985, P<0.001). Group B showed the best lip-closure morphology observed with naked eyes. Conclusions: When three output microelectrode of the AFNP stimulated side are separately imbedded into the upper lip, the lower lip and the connection with the corner of the mouth to the ipsilateral tragus, AFNP can sufficiently induce closed-mouth reaction. These positions are suitable as priority options microelectrodes implantation positions for the microelectrodes of the AFNP stimulated side.

Long-term follow-up of persistent pulmonary pure ground-glass nodules with deep learning-assisted nodule segmentation

OBJECTIVE

To investigate the natural history of persistent pulmonary pure ground-glass nodules (pGGNs) with deep learning-assisted nodule segmentation.

METHODS

Between January 2007 and October 2018, 110 pGGNs from 110 patients with 573 follow-up CT scans were included in this retrospective study. pGGN automatic segmentation was performed on initial and all follow-up CT scans using the Dr. Wise system based on convolution neural networks. Subsequently, pGGN diameter, density, volume, mass, volume doubling time (VDT), and mass doubling time (MDT) were calculated automatically. Enrolled pGGNs were categorized into growth, 52 (47.3%), and non-growth, 58 (52.7%), groups according to volume growth. Kaplan-Meier analyses with the log-rank test and Cox proportional hazards regression analysis were conducted to analyze the cumulative percentages of pGGN growth and identify risk factors for growth.

RESULTS

The mean follow-up period of the enrolled pGGNs was 48.7 ± 23.8 months. The median VDT of the 52 pGGNs having grown was 1448 (range, 339-8640) days, and their median MDT was 1332 (range, 290-38,912) days. The 12-month, 24.7-month, and 60.8-month cumulative percentages of pGGN growth were 10%, 25.5%, and 51.1%, respectively, and they significantly differed among the initial diameter, volume, and mass subgroups (all p < 0.001). The growth pattern of pGGNs may conform to the exponential model. Lobulated sign (p = 0.044), initial mean diameter (p < 0.001), volume (p = 0.003), and mass (p = 0.023) predicted pGGN growth.

CONCLUSIONS

Persistent pGGNs showed an indolent course. Deep learning can assist in accurately elucidating the natural history of pGGNs. pGGNs with lobulated sign and larger initial diameter, volume, and mass are more likely to grow.

KEY POINTS

• The pure ground-glass nodule (pGGN) segmentation accuracy of the Dr. Wise system based on convolution neural networks (CNNs) was 96.5% (573/594). • The median volume doubling time (VDT) of 52 pure ground-glass nodules (pGGNs) having grown was 1448 days (range, 339-8640 days), and their median mass doubling time (MDT) was 1332 days (range, 290-38,912 days). The mean time to growth in volume was 854 ± 675 days (range, 116-2856 days). • The 12-month, 24.7-month, and 60.8-month cumulative percentages of pGGN growth were 10%, 25.5%, and 51.1%, respectively, and they significantly differed among the initial diameter, volume, and mass subgroups (all p values < 0.001). The growth pattern of pure ground-glass nodules may conform to exponential model.

An Approach of Integrating Domain Knowledge into Data-Driven Diagnostic Model

A diagnostic model of general diseases could help general practitioners to decrease misdiagnoses and reduce workload. In this paper, we developed a neural network model that can classify potential diagnoses among 100 selected common diseases based on ambulatory health care data. We propose a novel approach to integrate domain knowledge into neural network training. The evaluation results show our model outperforming the baseline model in terms of knowledge consistency and model generalization.

Convolutional neural network-based segmentation can help in assessing the substantia nigra in neuromelanin MRI

Purpose

This study aimed to evaluate the accuracy and diagnostic test performance of the U-net-based segmentation method in neuromelanin magnetic resonance imaging (NM-MRI) compared to the established manual segmentation method for Parkinson’s disease (PD) diagnosis.

Methods

NM-MRI datasets from two different 3T-scanners were used: a “principal dataset” with 122 participants and an “external validation dataset” with 24 participants, including 62 and 12 PD patients, respectively. Two radiologists performed SNpc manual segmentation. Inter-reader precision was determined using Dice coefficients. The U-net was trained with manual segmentation as ground truth and Dice coefficients used to measure accuracy. Training and validation steps were performed on the principal dataset using a 4-fold cross-validation method. We tested the U-net on the external validation dataset. SNpc hyperintense areas were estimated from U-net and manual segmentation masks, replicating a previously validated thresholding method, and their diagnostic test performances for PD determined.

Results

For SNpc segmentation, U-net accuracy was comparable to inter-reader precision in the principal dataset (Dice coefficient: U-net, 0.83 ± 0.04; inter-reader, 0.83 ± 0.04), but lower in external validation dataset (Dice coefficient: U-net, 079 ± 0.04; inter-reader, 0.85 ± 0.03). Diagnostic test performances for PD were comparable between U-net and manual segmentation methods in both principal (area under the receiver operating characteristic curve: U-net, 0.950; manual, 0.948) and external (U-net, 0.944; manual, 0.931) datasets.

Conclusion

U-net segmentation provided relatively high accuracy in the evaluation of the SNpc in NM-MRI and yielded diagnostic performance comparable to that of the established manual method.

Electronic supplementary material

The online version of this article (10.1007/s00234-019-02279-w) contains supplementary material, which is available to authorized users.

Performance of machine learning software to classify breast lesions using BI-RADS radiomic features on ultrasound images

BACKGROUND

The purpose of this work was to evaluate computable Breast Imaging Reporting and Data System (BI-RADS) radiomic features to classify breast masses on ultrasound B-mode images.

METHODS

The database consisted of 206 consecutive lesions (144 benign and 62 malignant) proved by percutaneous biopsy in a prospective study approved by the local ethical committee. A radiologist manually delineated the contour of the lesions on greyscale images. We extracted the main ten radiomic features based on the BI-RADS lexicon and classified the lesions as benign or malignant using a bottom-up approach for five machine learning (ML) methods: multilayer perceptron (MLP), decision tree (DT), linear discriminant analysis (LDA), random forest (RF), and support vector machine (SVM). We performed a 10-fold cross validation for training and testing of all classifiers. Receiver operating characteristic (ROC) analysis was used for providing the area under the curve with 95% confidence intervals (CI).

RESULTS

The classifier with the highest AUC at ROC analysis was SVM (AUC = 0.840, 95% CI 0.6667-0.9762), with 71.4% sensitivity (95% CI 0.6479-0.8616) and 76.9% specificity (95% CI 0.6148-0.8228). The best AUC for each method was 0.744 (95% CI 0.677-0.774) for DT, 0.818 (95% CI 0.6667-0.9444) for LDA, 0.811 (95% CI 0.710-0.892) for RF, and 0.806 (95% CI 0.677-0.839) for MLP. Lesion margin and orientation were the optimal features for all the machine learning methods.

CONCLUSIONS

ML can aid the distinction between benign and malignant breast lesion on ultrasound images using quantified BI-RADS descriptors. SVM provided the highest ROC-AUC (0.840).

[Research progress in medical imaging based on deep learning of neural network]

The development of computer hardware allows rapid accumulation of medical imaging data. Deep learning has shown great potential in medical imaging data analysis and establish a new area of machine learning. The commonly used deep learning models were firstly introduced in the paper, and then, summarized with the application of deep learning in the detection, classification, diagnosis, segmentation, identification of medical imaging. The application of deep learning in oral and maxillofacial radiology and other discipline of stomatology was proposed. At the end, the paper discussed the problems of deep learning in medical imaging research.

Deep learning reconstruction improves image quality of abdominal ultra-high-resolution CT

OBJECTIVES

Deep learning reconstruction (DLR) is a new reconstruction method; it introduces deep convolutional neural networks into the reconstruction flow. This study was conducted in order to examine the clinical applicability of abdominal ultra-high-resolution CT (U-HRCT) exams reconstructed with a new DLR in comparison to hybrid and model-based iterative reconstruction (hybrid-IR, MBIR).

METHODS

Our retrospective study included 46 patients seen between December 2017 and April 2018. A radiologist recorded the standard deviation of attenuation in the paraspinal muscle as the image noise and calculated the contrast-to-noise ratio (CNR) for the aorta, portal vein, and liver. The overall image quality was assessed by two other radiologists and graded on a 5-point confidence scale ranging from 1 (unacceptable) to 5 (excellent). The difference between CT images subjected to hybrid-IR, MBIR, and DLR was compared.

RESULTS

The image noise was significantly lower and the CNR was significantly higher on DLR than hybrid-IR and MBIR images (p < 0.01). DLR images received the highest and MBIR images the lowest scores for overall image quality.

CONCLUSIONS

DLR improved the quality of abdominal U-HRCT images.

KEY POINTS

• The potential degradation due to increased noise may prevent implementation of ultra-high-resolution CT in the abdomen. • Image noise and overall image quality for hepatic ultra-high-resolution CT images improved with deep learning reconstruction as compared to hybrid- and model-based iterative reconstruction.

3D convolutional neural networks applied to CT angiography in the detection of acute ischemic stroke

Background

The aim of this study was to investigate the feasibility of ischemic stroke detection from computed tomography angiography source images (CTA-SI) using three-dimensional convolutional neural networks.

Methods

CTA-SI of 60 patients with a suspected acute ischemic stroke of the middle cerebral artery were randomly selected for this study; 30 patients were used in the neural network training, and the subsequent testing was performed using the remaining 30 patients. The training and testing were based on manually segmented lesions. Cerebral hemispheric comparison CTA and non-contrast computed tomography (NCCT) were studied as additional input features.

Results

All ischemic lesions in the testing data were correctly lateralized, and a high correspondence to manual segmentations was achieved. Patients with a diagnosed stroke had clinically relevant regions labeled infarcted with a 0.93 sensitivity and 0.82 specificity. The highest achieved voxel-wise area under receiver operating characteristic curve was 0.93, and the highest Dice similarity coefficient was 0.61. When cerebral hemispheric comparison was used as an input feature, the algorithm performance improved. Only a slight effect was seen when NCCT was included.

Conclusion

The results support the hypothesis that an acute ischemic stroke lesion can be detected with 3D convolutional neural network-based software from CTA-SI. Utilizing information from the contralateral hemisphere appears to be beneficial for reducing false positive findings.

Applications of machine learning algorithms to predict therapeutic outcomes in depression: A meta-analysis and systematic review

BACKGROUND

No previous study has comprehensively reviewed the application of machine learning algorithms in mood disorders populations. Herein, we qualitatively and quantitatively evaluate previous studies of machine learning-devised models that predict therapeutic outcomes in mood disorders populations.

METHODS

We searched Ovid MEDLINE/PubMed from inception to February 8, 2018 for relevant studies that included adults with bipolar or unipolar depression; assessed therapeutic outcomes with a pharmacological, neuromodulatory, or manual-based psychotherapeutic intervention for depression; applied a machine learning algorithm; and reported predictors of therapeutic response. A random-effects meta-analysis of proportions and meta-regression analyses were conducted.

RESULTS

We identified 639 records: 75 full-text publications were assessed for eligibility; 26 studies (n=17,499) and 20 studies (n=6325) were included in qualitative and quantitative review, respectively. Classification algorithms were able to predict therapeutic outcomes with an overall accuracy of 0.82 (95% confidence interval [CI] of [0.77, 0.87]). Pooled estimates of classification accuracy were significantly greater (p < 0.01) in models informed by multiple data types (e.g., composite of phenomenological patient features and neuroimaging or peripheral gene expression data; pooled proportion [95% CI] = 0.93[0.86, 0.97]) when compared to models with lower-dimension data types (pooledproportion=0.68[0.62,0.74]to0.85[0.81,0.88]).

LIMITATIONS

Most studies were retrospective; differences in machine learning algorithms and their implementation (e.g., cross-validation, hyperparameter tuning); cannot infer importance of individual variables fed into learning algorithm.

CONCLUSIONS

Machine learning algorithms provide a powerful conceptual and analytic framework capable of integrating multiple data types and sources. An integrative approach may more effectively model neurobiological components as functional modules of pathophysiology embedded within the complex, social dynamics that influence the phenomenology of mental disorders.

Artificial intelligence in medical imaging: threat or opportunity? Radiologists again at the forefront of innovation in medicine

One of the most promising areas of health innovation is the application of artificial intelligence (AI), primarily in medical imaging. This article provides basic definitions of terms such as “machine/deep learning” and analyses the integration of AI into radiology. Publications on AI have drastically increased from about 100–150 per year in 2007–2008 to 700–800 per year in 2016–2017. Magnetic resonance imaging and computed tomography collectively account for more than 50% of current articles. Neuroradiology appears in about one-third of the papers, followed by musculoskeletal, cardiovascular, breast, urogenital, lung/thorax, and abdomen, each representing 6–9% of articles. With an irreversible increase in the amount of data and the possibility to use AI to identify findings either detectable or not by the human eye, radiology is now moving from a subjective perceptual skill to a more objective science. Radiologists, who were on the forefront of the digital era in medicine, can guide the introduction of AI into healthcare. Yet, they will not be replaced because radiology includes communication of diagnosis, consideration of patient’s values and preferences, medical judgment, quality assurance, education, policy-making, and interventional procedures. The higher efficiency provided by AI will allow radiologists to perform more value-added tasks, becoming more visible to patients and playing a vital role in multidisciplinary clinical teams.

Application of artificial intelligence using a convolutional neural network for detecting gastric cancer in endoscopic images

BACKGROUND

Image recognition using artificial intelligence with deep learning through convolutional neural networks (CNNs) has dramatically improved and been increasingly applied to medical fields for diagnostic imaging. We developed a CNN that can automatically detect gastric cancer in endoscopic images.

METHODS

A CNN-based diagnostic system was constructed based on Single Shot MultiBox Detector architecture and trained using 13,584 endoscopic images of gastric cancer. To evaluate the diagnostic accuracy, an independent test set of 2296 stomach images collected from 69 consecutive patients with 77 gastric cancer lesions was applied to the constructed CNN.

RESULTS

The CNN required 47 s to analyze 2296 test images. The CNN correctly diagnosed 71 of 77 gastric cancer lesions with an overall sensitivity of 92.2%, and 161 non-cancerous lesions were detected as gastric cancer, resulting in a positive predictive value of 30.6%. Seventy of the 71 lesions (98.6%) with a diameter of 6 mm or more as well as all invasive cancers were correctly detected. All missed lesions were superficially depressed and differentiated-type intramucosal cancers that were difficult to distinguish from gastritis even for experienced endoscopists. Nearly half of the false-positive lesions were gastritis with changes in color tone or an irregular mucosal surface.

CONCLUSION

The constructed CNN system for detecting gastric cancer could process numerous stored endoscopic images in a very short time with a clinically relevant diagnostic ability. It may be well applicable to daily clinical practice to reduce the burden of endoscopists.

Neural Network Classifier for Automatic Detection of Invasive Versus Noninvasive Airway Management Technique Based on Respiratory Monitoring Parameters in a Pediatric Anesthesia

Children undergoing general anesthesia require airway monitoring by an anesthesia provider. The airway may be supported with noninvasive devices such as face mask or invasive devices such as a laryngeal mask airway or an endotracheal tube. The physiologic data stored provides an opportunity to apply machine learning algorithms distinguish between these modes based on pattern recognition. We retrieved three data sets from patients receiving general anesthesia in 2015 with either mask, laryngeal mask airway or endotracheal tube. Patients underwent myringotomy, tonsillectomy, adenoidectomy or inguinal hernia repair procedures. We retrieved measurements for end-tidal carbon dioxide, tidal volume, and peak inspiratory pressure and calculated statistical features for each data element per patient. We applied machine learning algorithms (decision tree, support vector machine, and neural network) to classify patients into noninvasive or invasive airway device support. We identified 300 patients per group (mask, laryngeal mask airway, and endotracheal tube) for a total of 900 patients. The neural network classifier performed better than the boosted trees and support vector machine classifiers based on the test data sets. The sensitivity, specificity, and accuracy for neural network classification are 97.5%, 96.3%, and 95.8%. In contrast, the sensitivity, specificity, and accuracy of support vector machine are 89.1%, 92.3%, and 88.3% and with the boosted tree classifier they are 93.8%, 92.1%, and 91.4%. We describe a method to automatically distinguish between noninvasive and invasive airway device support in a pediatric surgical setting based on respiratory monitoring parameters. The results show that the neural network classifier algorithm can accurately classify noninvasive and invasive airway device support.

[The application of artificial neural network on the assessment of lexical tone production of pediatric cochlear implant users]

Objective: The present study was carried out to explore the tone production ability of the Mandarin-speaking children with cochlear implants (CI) by using an artificial neural network model and to examine the potential contributing factors underlining their tone production performance. The results of this study might provide useful guidelines for post-operative rehabilitation processes of pediatric CI users. Methods: Two hundred and seventy-eight prelingually deafened children who received unilateral CI participated in this study. As controls, 170 similarly-aged children with normal hearing (NH) were recruited. A total of 36 Chinese monosyllabic words were selected as the tone production targets. Vocal production samples were recorded and the fundamental frequency (F0) contour of each syllable was extracted using an auto-correlation algorithm followed by manual correction. An artificial neural network was created in MATLAB to classify the tone production. The relationships between tone production and several demographic factors were evaluated. Results: Pediatric CI users produced Mandarin tones much less accurately than did the NH children (58.8% vs. 91.5% correct). Tremendous variability in tone production performance existed among the CI children. Tones 2 and 3 were produced less accurately than tones 1 and 4 for both groups. For the CI group, all tones when in error tended to be judged as tone 1. The tone production accuracy was negatively correlated with age at implantation and positively correlated with CI use duration with correlation coefficients (r) of -0.215 (P=0.003) and 0.203 (P=0.005), respectively. Age was one of the determinants of tonal ability for NH children. Conclusions: For children with severe to profound hearing loss, early implantation and persistent use of CI are beneficial to their tone production development. Artificial neural network is a convenient and reliable assessment tool for the development of tonal ability of hearing-impaired children who are in the rehabilitation processes that focus on speech and language expression.

Artificial Neural Network Approach in Laboratory Test Reporting: Learning Algorithms

OBJECTIVES

In the field of laboratory medicine, minimizing errors and establishing standardization is only possible by predefined processes. The aim of this study was to build an experimental decision algorithm model open to improvement that would efficiently and rapidly evaluate the results of biochemical tests with critical values by evaluating multiple factors concurrently.

METHODS

The experimental model was built by Weka software (Weka, Waikato, New Zealand) based on the artificial neural network method. Data were received from Dokuz Eylül University Central Laboratory. "Training sets" were developed for our experimental model to teach the evaluation criteria. After training the system, "test sets" developed for different conditions were used to statistically assess the validity of the model.

RESULTS

After developing the decision algorithm with three iterations of training, no result was verified that was refused by the laboratory specialist. The sensitivity of the model was 91% and specificity was 100%. The estimated κ score was 0.950.

CONCLUSIONS

This is the first study based on an artificial neural network to build an experimental assessment and decision algorithm model. By integrating our trained algorithm model into a laboratory information system, it may be possible to reduce employees' workload without compromising patient safety.