A Postreconstruction Harmonization Method for Multicenter Radiomic Studies in PET

Combatting the effect of image reconstruction settings on lymphoma [18F]FDG PET metabolic tumor volume assessment using various segmentation methods

Background

[¹⁸F]FDG PET-based metabolic tumor volume (MTV) is a promising prognostic marker for lymphoma patients. The aim of this study is to assess the sensitivity of several MTV segmentation methods to variations in image reconstruction methods and the ability of ComBat to improve MTV reproducibility.

Methods

Fifty-six lesions were segmented from baseline [¹⁸F]FDG PET scans of 19 lymphoma patients. For each scan, EARL1 and EARL2 standards and locally clinically preferred reconstruction protocols were applied. Lesions were delineated using 9 semiautomatic segmentation methods: fixed threshold based on standardized uptake value (SUV), (SUV = 4, SUV = 2.5), relative threshold (41% of SUVmax [41M], 50% of SUVpeak [A50P]), majority vote-based methods that select voxels detected by at least 2 (MV2) and 3 (MV3) out of the latter 4 methods, Nestle thresholding, and methods that identify the optimal method based on SUVmax (L2A, L2B). MTVs from EARL2 and locally clinically preferred reconstructions were compared to those from EARL1. Finally, different versions of ComBat were explored to harmonize the data.

Results

MTVs from the SUV4.0 method were least sensitive to the use of different reconstructions (MTV ratio: median = 1.01, interquartile range = [0.96–1.10]). After ComBat harmonization, an improved agreement of MTVs among different reconstructions was found for most segmentation methods. The regular implementation of ComBat (‘Regular ComBat’) using non-transformed distributions resulted in less accurate and precise MTV alignments than a version using log-transformed datasets (‘Log-transformed ComBat’).

Conclusion

MTV depends on both segmentation method and reconstruction methods. ComBat reduces reconstruction dependent MTV variability, especially when log-transformation is used to account for the non-normal distribution of MTVs.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13550-022-00916-9.

Self-supervised learning for multi-center magnetic resonance imaging harmonization without traveling phantoms

Objective. With the progress of artificial intelligence (AI) in magnetic resonance imaging (MRI), large-scale multi-center MRI datasets have a great influence on diagnosis accuracy and model performance. However, multi-center images are highly variable due to the variety of scanners or scanning parameters in use, which has a negative effect on the generality of AI-based diagnosis models. To address this problem, we propose a self-supervised harmonization (SSH) method. Approach. Mapping the style of images between centers allows harmonization without traveling phantoms to be formalized as an unpaired image-to-image translation problem between two domains. The mapping is a two-stage transform, consisting of a modified cycle generative adversarial network (cycleGAN) for style transfer and a histogram matching module for structure fidelity. The proposed algorithm is demonstrated using female pelvic MRI images from two 3 T systems and compared with three state-of-the-art methods and one conventional method. In the absence of traveling phantoms, we evaluate harmonization from three perspectives: image fidelity, ability to remove inter-center differences, and influence on the downstream model. Main results. The improved image sharpness and structure fidelity are observed using the proposed harmonization pipeline. It largely decreases the number of features with a significant difference between two systems (from 64 to 45, lower than dualGAN: 57, cycleGAN: 59, ComBat: 64, and CLAHE: 54). In the downstream cervical cancer classification, it yields an area under the receiver operating characteristic curve of 0.894 (higher than dualGAN: 0.828, cycleGAN: 0.812, ComBat: 0.685, and CLAHE: 0.770). Significance. Our SSH method yields superior generality of downstream cervical cancer classification models by significantly decreasing the difference in radiomics features, and it achieves greater image fidelity.

Artificial intelligence technologies in nuclear medicine

The use of artificial intelligence plays a crucial role in developing precision medicine in nuclear medicine. Artificial intelligence refers to a field of computer science aimed at imitating the performance of tasks typically requiring human intelligence. From machine learning to generative adversarial networks, artificial intelligence automized the workflow of medical imaging. In this mini-review, we encapsulate artificial intelligence models and their use in nuclear medicine imaging workflow.

Quantitative classification and radiomics of [18F]FDG-PET/CT in indeterminate thyroid nodules

PURPOSE

To evaluate whether quantitative [18F]FDG-PET/CT assessment, including radiomic analysis of [18F]FDG-positive thyroid nodules, improved the preoperative differentiation of indeterminate thyroid nodules of non-Hürthle cell and Hürthle cell cytology.

METHODS

Prospectively included patients with a Bethesda III or IV thyroid nodule underwent [18F]FDG-PET/CT imaging. Receiver operating characteristic (ROC) curve analysis was performed for standardised uptake values (SUV) and SUV-ratios, including assessment of SUV cut-offs at which a malignant/borderline neoplasm was reliably ruled out (≥ 95% sensitivity). [18F]FDG-positive scans were included in radiomic analysis. After segmentation at 50% of SUVpeak, 107 radiomic features were extracted from [18F]FDG-PET and low-dose CT images. Elastic net regression classifiers were trained in a 20-times repeated random split. Dimensionality reduction was incorporated into the splits. Predictive performance of radiomics was presented as mean area under the ROC curve (AUC) across the test sets.

RESULTS

Of 123 included patients, 84 (68%) index nodules were visually [18F]FDG-positive. The malignant/borderline rate was 27% (33/123). SUV-metrices showed AUCs ranging from 0.705 (95% CI, 0.601-0.810) to 0.729 (0.633-0.824), 0.708 (0.580-0.835) to 0.757 (0.650-0.864), and 0.533 (0.320-0.747) to 0.700 (0.502-0.898) in all (n = 123), non-Hürthle (n = 94), and Hürthle cell (n = 29) nodules, respectively. At SUVmax, SUVpeak, SUVmax-ratio, and SUVpeak-ratio cut-offs of 2.1 g/mL, 1.6 g/mL, 1.2, and 0.9, respectively, sensitivity of [18F]FDG-PET/CT was 95.8% (95% CI, 78.9-99.9%) in non-Hürthle cell nodules. In Hürthle cell nodules, cut-offs of 5.2 g/mL, 4.7 g/mL, 3.4, and 2.8, respectively, resulted in 100% sensitivity (95% CI, 66.4-100%). Radiomic analysis of 84 (68%) [18F]FDG-positive nodules showed a mean test set AUC of 0.445 (95% CI, 0.290-0.600) for the PET model.

CONCLUSION

Quantitative [18F]FDG-PET/CT assessment ruled out malignancy in indeterminate thyroid nodules. Distinctive, higher SUV cut-offs should be applied in Hürthle cell nodules to optimize rule-out ability. Radiomic analysis did not contribute to the additional differentiation of [18F]FDG-positive nodules.

TRIAL REGISTRATION NUMBER

This trial is registered with ClinicalTrials.gov: NCT02208544 (5 August 2014), https://clinicaltrials.gov/ct2/show/NCT02208544 .

Voxel-wise supervised analysis of tumors with multimodal engineered features to highlight interpretable biological patterns

BACKGROUND

Translation of predictive and prognostic image-based learning models to clinical applications is challenging due in part to their lack of interpretability. Some deep-learning-based methods provide information about the regions driving the model output. Yet, due to the high-level abstraction of deep features, these methods do not completely solve the interpretation challenge. In addition, low sample size cohorts can lead to instabilities and suboptimal convergence for models involving a large number of parameters such as convolutional neural networks.

PURPOSE

Here, we propose a method for designing radiomic models that combines the interpretability of handcrafted radiomics with a sub-regional analysis.

MATERIALS AND METHODS

Our approach relies on voxel-wise engineered radiomic features with average global aggregation and logistic regression. The method is illustrated using a small dataset of 51 soft tissue sarcoma (STS) patients where the task is to predict the risk of lung metastasis occurrence during the follow-up period.

RESULTS

Using positron emission tomography/computed tomography and two magnetic resonance imaging sequences separately to build two radiomic models, we show that our approach produces quantitative maps that highlight the signal that contributes to the decision within the tumor region of interest. In our STS example, the analysis of these maps identified two biological patterns that are consistent with STS grading systems and knowledge: necrosis development and glucose metabolism of the tumor.

CONCLUSIONS

We demonstrate how that method makes it possible to spatially and quantitatively interpret radiomic models amenable to sub-regions identification and biological interpretation for patient stratification.

CT-Based Radiomics and Deep Learning for BRCA Mutation and Progression-Free Survival Prediction in Ovarian Cancer Using a Multicentric Dataset

PURPOSE

Build predictive radiomic models for early relapse and BRCA mutation based on a multicentric database of high-grade serous ovarian cancer (HGSOC) and validate them in a test set coming from different institutions.

METHODS

Preoperative CTs of patients with HGSOC treated at four referral centers were retrospectively acquired and manually segmented. Hand-crafted features and deep radiomics features were extracted respectively by dedicated software (MODDICOM) and a dedicated convolutional neural network (CNN). Features were selected with and without prior harmonization (ComBat harmonization), and models were built using different machine learning algorithms, including clinical variables.

RESULTS

We included 218 patients. Radiomic models showed low performance in predicting both BRCA mutation (AUC in test set between 0.46 and 0.59) and 1-year relapse (AUC in test set between 0.46 and 0.56); deep learning models demonstrated similar results (AUC in the test of 0.48 for BRCA and 0.50 for relapse). The inclusion of clinical variables improved the performance of the radiomic models to predict BRCA mutation (AUC in the test set of 0.74).

CONCLUSIONS

In our multicentric dataset, representative of a real-life clinical scenario, we could not find a good radiomic predicting model for PFS and BRCA mutational status, with both traditional radiomics and deep learning, but the combination of clinical and radiomic models improved model performance for the prediction of BRCA mutation. These findings highlight the need for standardization through the whole radiomic pipelines and robust multicentric external validations of results.

Harmonization of multi-site MRS data with ComBat

Magnetic resonance spectroscopy (MRS) is a non-invasive neuroimaging technique used to measure brain chemistry in vivo and has been used to study the healthy brain as well as neuropathology in numerous neurological disorders. The number of multi-site studies using MRS are increasing; however, non-biological variability introduced during data collection across multiple sites, such as differences in scanner vendors and site-specific acquisition implementations for MRS, can obscure detection of biological effects of interest. ComBat is a data harmonization technique that can remove non-biological sources of variance in multisite studies. It has been validated for use with structural and functional MRI metrics but not for MRS metabolites. This study investigated the validity of using ComBat to harmonize MRS metabolites for vendor and site differences. Analyses were performed using data acquired across 20 sites and included edited MRS for GABA+ (N=218) and macromolecule-suppressed GABA data (N=209), as well as standard PRESS data to quantify NAA, creatine, choline, and glutamate (N=190). ComBat harmonization successfully mitigated vendor and site differences for all metabolites of interest. Moreover, significant associations were detected between sex and choline levels and between age and glutamate and GABA+ levels that were not detectable prior to harmonization, confirming the importance of removing site and vendor effects in multi-site data. In conclusion, ComBat harmonization can be successfully applied to MRS data in multi-site MRS studies.

A tri-modal tissue-equivalent anthropomorphic phantom for PET, CT and multi-parametric MRI radiomics

PURPOSE

Radiomics has emerged as an advanced image processing methodology to define quantitative imaging biomarkers for prognosis and prediction of treatment response and outcome. The development of quantitative imaging biomarkers requires careful analysis to define their accuracy, stability and reproducibility through phantom measurements. Few efforts were devoted to develop realistic anthropomorphic phantoms. In this work, we developed a multimodality image phantom suitable for PET, CT and multiparametric MRI imaging.

METHODS

A tissue-equivalent gel-based mixture was designed and tested for compatibility with different imaging modalities. Calibration measurements allowed to assess gel composition to simulate PET, CT and MRI contrasts of oncological lesions. The characterized gel mixture was used to create realistic synthetic lesions (e.g. lesions with irregular shape and non-uniform image contrast), to be inserted in a standard anthropomorphic phantom. In order to show phantom usefulness, issues related to accuracy, stability and reproducibility of radiomic biomarkers were addressed as proofs-of-concept.

RESULTS

The procedure for gel preparation was straightforward and the characterized gel mixture allowed to mime simultaneously oncological lesion contrast in CT, PET and MRI imaging. Proofs-of-concept studies suggested that phantom measurements can be customized for specific clinical situations and radiomic protocols.

CONCLUSIONS

We developed a strategy to manufacture an anthropomorphic, tissue-equivalent, multimodal phantom to be customized on specific radiomics protocols, for addressing specific methodological issues both in mono and multicentric studies.

Noise-Based Image Harmonization Significantly Increases Repeatability and Reproducibility of Radiomics Features in PET Images: A Phantom Study

For multicenter clinical studies, characterizing the robustness of image-derived radiomics features is essential. Features calculated on PET images have been shown to be very sensitive to image noise. The purpose of this work was to investigate the efficacy of a relatively simple harmonization strategy on feature robustness and agreement. A purpose-built texture pattern phantom was scanned on 10 different PET scanners in 7 institutions with various different image acquisition and reconstruction protocols. An image harmonization technique based on equalizing a contrast-to-noise ratio was employed to generate a "harmonized" alongside a "standard" dataset for a reproducibility study. In addition, a repeatability study was performed with images from a single PET scanner of variable image noise, varying the binning time of the reconstruction. Feature agreement was measured using the intraclass correlation coefficient (ICC). In the repeatability study, 81/93 features had a lower ICC on the images with the highest image noise as compared to the images with the lowest image noise. Using the harmonized dataset significantly improved the feature agreement for five of the six investigated feature classes over the standard dataset. For three feature classes, high feature agreement corresponded with higher sensitivity to the different patterns, suggesting a way to select suitable features for predictive models.

Repeatability and Reproducibility of Computed Tomography Radiomics for Pulmonary Nodules: A Multicenter Phantom Study

BACKGROUND

Radiomics can yield minable information from medical images, which can facilitate computer-aided diagnosis. However, the lack of repeatability and reproducibility of radiomic features (RFs) may hinder their generalizability in clinical applications.

OBJECTIVES

The aims of this study were to explore 3 main sources of variability in RFs, investigate their influencing magnitudes and patterns, and identify a subset of robust RFs for further studies.

MATERIALS AND METHODS

A chest phantom with nodules was scanned with different computed tomography (CT) scanners repeatedly with varying acquisition and reconstruction parameters (April-May 2019) to evaluate 3 sources of variability: test-retest, inter-CT, and intra-CT protocol variability. The robustness of the RFs was measured using the concordance correlation coefficient, dynamic range, and intraclass correlation coefficient (ICC). The influencing magnitudes and patterns were analyzed using the Friedman test and Spearman rank correlation coefficient. Stable and informative RFs were selected, and their redundancy was eliminated using hierarchical clustering. Clinical validation was also performed to verify the clinical effectiveness and potential enhancement of the generalizability of radiomics research.

RESULTS

A total of 1295 RFs that showed all 3 sources of variability were included. The reconstruction kernel and the iteration level showed the greatest (ICC, 0.35 ± 0.31) and the least (ICC, 0.63 ± 0.27) influence on magnitudes. The different sources of variability showed relatively consistent patterns of influence (false discovery rate <0.001). Finally, we obtained a subset of 19 stable, informative, and nonredundant RFs under all 3 sources of variability. These RFs exhibited clinical effectiveness and showed better prediction performance than unstable RFs in the validation dataset (P = 0.017, Delong test).

CONCLUSIONS

The stability of RFs was affected to different degrees by test-retest and differences in CT manufacturers and models and CT acquisition and reconstruction parameters, but the influences of these factors showed relatively consistent patterns. We also obtained a subset of 19 stable, informative, and nonredundant RFs that should be preferably used to enhance the generalizability of further radiomics research.

Context-Aware Saliency Guided Radiomics: Application to Prediction of Outcome and HPV-Status from Multi-Center PET/CT Images of Head and Neck Cancer

Simple Summary

This study investigated the ability of context-aware saliency-guided PET/CT radiomics in the prediction of outcome and HPV status for head and neck cancer. In total, 806 HNC patients (training vs. validation vs. external testing: 500 vs. 97 vs. 209) from 9 centers were collected from The Cancer Imaging Archive (TCIA). Saliency-guided radiomics showed enhanced performance for both outcome and HPV-status predictions relative to conventional radiomics. The radiomics-predicted HPV status also showed complementary prognostic value. This multi-center study highlights the feasibility of saliency-guided PET/CT radiomics in outcome predictions of head and neck cancer, confirming that certain regions are more relevant to tumor aggressiveness and prognosis.

Abstract

Purpose: This multi-center study aims to investigate the prognostic value of context-aware saliency-guided radiomics in ¹⁸F-FDG PET/CT images of head and neck cancer (HNC). Methods: 806 HNC patients (training vs. validation vs. external testing: 500 vs. 97 vs. 209) from 9 centers were collected from The Cancer Imaging Archive (TCIA). There were 100/384 and 60/123 oropharyngeal carcinoma (OPC) patients with human papillomavirus (HPV) status in training and testing cohorts, respectively. Six types of images were used for radiomics feature extraction and further model construction, namely (i) the original image (Origin), (ii) a context-aware saliency map (SalMap), (iii, iv) high- or low-saliency regions in the original image (highSal or lowSal), (v) a saliency-weighted image (SalxImg), and finally, (vi) a fused PET-CT image (FusedImg). Four outcomes were evaluated, i.e., recurrence-free survival (RFS), metastasis-free survival (MFS), overall survival (OS), and disease-free survival (DFS), respectively. Multivariate Cox analysis and logistic regression were adopted to construct radiomics scores for the prediction of outcome (Rad_Ocm) and HPV-status (Rad_HPV), respectively. Besides, the prognostic value of their integration (Rad_Ocm_HPV) was also investigated. Results: In the external testing cohort, compared with the Origin model, SalMap and SalxImg achieved the highest C-indices for RFS (0.621 vs. 0.559) and MFS (0.785 vs. 0.739) predictions, respectively, while FusedImg performed the best for both OS (0.685 vs. 0.659) and DFS (0.641 vs. 0.582) predictions. In the OPC HPV testing cohort, FusedImg showed higher AUC for HPV-status prediction compared with the Origin model (0.653 vs. 0.484). In the OPC testing cohort, compared with Rad_Ocm or Rad_HPV alone, Rad_Ocm_HPV performed the best for OS and DFS predictions with C-indices of 0.702 (p = 0.002) and 0.684 (p = 0.006), respectively. Conclusion: Saliency-guided radiomics showed enhanced performance for both outcome and HPV-status predictions relative to conventional radiomics. The radiomics-predicted HPV status also showed complementary prognostic value.

Reproducibility of Standardized Uptake Values Including Volume Metrics Between TOF-PET-MR and TOF-PET-CT

Purpose

To investigate the reproducibility of tracer uptake measurements, including volume metrics, such as metabolic tumor volume (MTV) and tumor lesion glycolysis (TLG) obtained by TOF-PET-CT and TOF-PET-MR.

Materials and Methods

Eighty consecutive patients with different oncologic diagnoses underwent TOF-PET-CT (Discovery 690; GE Healthcare) and TOF-PET-MR (SIGNA PET-MR; GE Healthcare) on the same day with single dose−18F-FDG injection. The scan order, PET-CT following or followed by PET-MR, was randomly assigned. A spherical volume of interest (VOI) of 30 mm was placed on the liver in accordance with the PERCIST criteria. For liver, the maximum and mean standard uptake value for body weight (SUV) and lean body mass (SUL) were obtained. For tumor delineation, VOI with a threshold of 40 and 50% of SUVmax was used (VOI40 and VOI50). The SUVmax, SUVmean, SUVpeak, MTV and TLG were calculated. The measurements were compared between the two scanners.

Results

In total, 80 tumor lesions from 35 patients were evaluated. There was no statistical difference observed in liver regions, whereas in tumor lesions, SUVmax, SUV mean, and SUVpeak of PET-MR were significantly underestimated (p < 0.001) in both VOI40 and VOI50. Among volume metrics, there was no statistical difference observed except TLG on VOI50 (p = 0.03). Correlation between PET-CT and PET-MR of each metrics were calculated. There was a moderate correlation of the liver SUV and SUL metrics (r = 0.63–0.78). In tumor lesions, SUVmax and SUVmean had a stronger correlation with underestimation in PET-MR on VOI 40 (SUVmax and SUVmean; r = 0.92 and 0.91 with slope = 0.71 and 0.72, respectively). In the evaluation of MTV and TLG, the stronger correlations were observed both on VOI40 (MTV and TLG; r = 0.75 and 0.92) and VOI50 (MTV and TLG; r = 0.88 and 0.95) between PET-CT and PET-MR.

Conclusion

PET metrics on TOF-PET-MR showed a good correlation with that of TOF-PET-CT. SUVmax and SUVpeak of tumor lesions were underestimated by 16% on PET-MRI. MTV with % threshold can be regarded as identical volumetric markers for both TOF-PET-CT and TOF-PET-MR.

Segmentation Uncertainty Estimation as a Sanity Check for Image Biomarker Studies

Problem. Image biomarker analysis, also known as radiomics, is a tool for tissue characterization and treatment prognosis that relies on routinely acquired clinical images and delineations. Due to the uncertainty in image acquisition, processing, and segmentation (delineation) protocols, radiomics often lack reproducibility. Radiomics harmonization techniques have been proposed as a solution to reduce these sources of uncertainty and/or their influence on the prognostic model performance. A relevant question is how to estimate the protocol-induced uncertainty of a specific image biomarker, what the effect is on the model performance, and how to optimize the model given the uncertainty. Methods. Two non-small cell lung cancer (NSCLC) cohorts, composed of 421 and 240 patients, respectively, were used for training and testing. Per patient, a Monte Carlo algorithm was used to generate three hundred synthetic contours with a surface dice tolerance measure of less than 1.18 mm with respect to the original GTV. These contours were subsequently used to derive 104 radiomic features, which were ranked on their relative sensitivity to contour perturbation, expressed in the parameter η. The top four (low η) and the bottom four (high η) features were selected for two models based on the Cox proportional hazards model. To investigate the influence of segmentation uncertainty on the prognostic model, we trained and tested the setup in 5000 augmented realizations (using a Monte Carlo sampling method); the log-rank test was used to assess the stratification performance and stability of segmentation uncertainty. Results. Although both low and high η setup showed significant testing set log-rank p-values (p = 0.01) in the original GTV delineations (without segmentation uncertainty introduced), in the model with high uncertainty, to effect ratio, only around 30% of the augmented realizations resulted in model performance with p < 0.05 in the test set. In contrast, the low η setup performed with a log-rank p < 0.05 in 90% of the augmented realizations. Moreover, the high η setup classification was uncertain in its predictions for 50% of the subjects in the testing set (for 80% agreement rate), whereas the low η setup was uncertain only in 10% of the cases. Discussion. Estimating image biomarker model performance based only on the original GTV segmentation, without considering segmentation, uncertainty may be deceiving. The model might result in a significant stratification performance, but can be unstable for delineation variations, which are inherent to manual segmentation. Simulating segmentation uncertainty using the method described allows for more stable image biomarker estimation, selection, and model development. The segmentation uncertainty estimation method described here is universal and can be extended to estimate other protocol uncertainties (such as image acquisition and pre-processing).

Overview of radiomics in prostate imaging and future directions

Recent advancements in imaging technology and analysis methods have led to an analytic framework known as radiomics. This framework extracts comprehensive high-dimensional features from imaging data and performs data mining to build analytical models for improved decision-support. Its features include many categories spanning texture and shape; thus, it can provide abundant information for precision medicine. Many studies of prostate radiomics have shown promising results in the assessment of pathological features, prediction of treatment response, and stratification of risk groups. Herein, we aimed to provide a general overview of radiomics procedures, discuss technical issues, explain various clinical applications, and suggest future research directions, especially for prostate imaging.

Outcome Prediction at Patient Level Derived from Pre-Treatment 18F-FDG PET Due to Machine Learning in Metastatic Melanoma Treated with Anti-PD1 Treatment

(1) Background: As outcome of patients with metastatic melanoma treated with anti-PD1 immunotherapy can vary in success, predictors are needed. We aimed to predict at the patients’ levels, overall survival (OS) and progression-free survival (PFS) after one year of immunotherapy, based on their pre-treatment 18F-FDG PET; (2) Methods: Fifty-six metastatic melanoma patients—without prior systemic treatment—were retrospectively included. Forty-five 18F-FDG PET-based radiomic features were computed and the top five features associated with the patient’s outcome were selected. The analyzed machine learning classifiers were random forest (RF), neural network, naive Bayes, logistic regression and support vector machine. The receiver operating characteristic curve was used to compare model performances, which were validated by cross-validation; (3) Results: The RF model obtained the best performance after validation to predict OS and PFS and presented AUC, sensitivities and specificities (IC95%) of 0.87 ± 0.1, 0.79 ± 0.11 and 0.95 ± 0.06 for OS and 0.9 ± 0.07, 0.88 ± 0.09 and 0.91 ± 0.08 for PFS, respectively. (4) Conclusion: A RF classifier, based on pretreatment 18F-FDG PET radiomic features may be useful for predicting the survival status for melanoma patients, after one year of a first line systemic treatment by immunotherapy.

Machine learning based evaluation of clinical and pretreatment 18F-FDG-PET/CT radiomic features to predict prognosis of cervical cancer patients

PURPOSE

To examine the usefulness of machine learning to predict prognosis in cervical cancer using clinical and radiomic features of 2-deoxy-2-[¹⁸F]fluoro-D-glucose (¹⁸F-FDG) positron emission tomography/computed tomography (CT) (¹⁸F-FDG-PET/CT).

METHODS

This retrospective study included 50 cervical cancer patients who underwent ¹⁸F-FDG-PET/CT before treatment. Four clinical (age, histology, stage, and treatment) and 41 ¹⁸F-FDG-PET-based radiomic features were ranked and a subset of useful features for association with disease progression was selected based on decrease of the Gini impurity. Six machine learning algorithms (random forest, neural network, k-nearest neighbors, naive Bayes, logistic regression, and support vector machine) were compared using the areas under the receiver operating characteristic curve (AUC). Progression-free survival (PFS) was assessed using Cox regression analysis.

RESULTS

The five top predictors of disease progression were: stage, surface area, metabolic tumor volume, gray-level run length non-uniformity (GLRLM_RLNU), and gray-level non-uniformity for run (GLRLM_GLNU). The naive Bayes model was the best-performing classifier for predicting disease progression (AUC = 0.872, accuracy = 0.780, F1 score = 0.781, precision = 0.788, and recall = 0.780). In the naive Bayes model, 5-year PFS was significantly higher in predicted non-progression than predicted progression (80.1% vs. 9.1%, p < 0.001) and was only the independent factor for PFS in multivariate analysis (HR, 6.89; 95% CI, 1.92-24.69; p = 0.003).

CONCLUSION

A machine learning approach based on clinical and pretreatment ¹⁸F-FDG PET-based radiomic features may be useful for predicting tumor progression in cervical cancer patients.

Differential Diagnosis and Molecular Stratification of Gastrointestinal Stromal Tumors on CT Images Using a Radiomics Approach

Treatment planning of gastrointestinal stromal tumors (GISTs) includes distinguishing GISTs from other intra-abdominal tumors and GISTs’ molecular analysis. The aim of this study was to evaluate radiomics for distinguishing GISTs from other intra-abdominal tumors, and in GISTs, predict the c-KIT, PDGFRA, BRAF mutational status, and mitotic index (MI). Patients diagnosed at the Erasmus MC between 2004 and 2017, with GIST or non-GIST intra-abdominal tumors and a contrast-enhanced venous-phase CT, were retrospectively included. Tumors were segmented, from which 564 image features were extracted. Prediction models were constructed using a combination of machine learning approaches. The evaluation was performed in a 100 × random-split cross-validation. Model performance was compared to that of three radiologists. One hundred twenty-five GISTs and 122 non-GISTs were included. The GIST vs. non-GIST radiomics model had a mean area under the curve (AUC) of 0.77. Three radiologists had an AUC of 0.69, 0.76, and 0.84, respectively. The radiomics model had an AUC of 0.52 for c-KIT, 0.56 for c-KIT exon 11, and 0.52 for the MI. The numbers of PDGFRA, BRAF, and other c-KIT mutations were too low for analysis. Our radiomics model was able to distinguish GISTs from non-GISTs with a performance similar to three radiologists, but less observer dependent. Therefore, it may aid in the early diagnosis of GIST, facilitating rapid referral to specialized treatment centers. As the model was not able to predict any genetic or molecular features, it cannot aid in treatment planning yet.

Reproducibility of radiomic features in CT images of NSCLC patients: an integrative analysis on the impact of acquisition and reconstruction parameters

BACKGROUND

We investigated to what extent tube voltage, scanner model, and reconstruction algorithm affect radiomic feature reproducibility in a single-institution retrospective database of computed tomography images of non-small-cell lung cancer patients.

METHODS

This study was approved by the Institutional Review Board (UID 2412). Images of 103 patients were considered, being acquired on either among two scanners, at 100 or 120 kVp. For each patient, images were reconstructed with six iterative blending levels, and 1414 features were extracted from each reconstruction. At univariate analysis, Wilcoxon-Mann-Whitney test was applied to evaluate feature differences within scanners and voltages, whereas the impact of the reconstruction was established with the overall concordance correlation coefficient (OCCC). A multivariable mixed model was also applied to investigate the independent contribution of each acquisition/reconstruction parameter. Univariate and multivariable analyses were combined to analyse feature behaviour.

RESULTS

Scanner model and voltage did not affect features significantly. The reconstruction blending level showed a significant impact at both univariate analysis (154/1414 features yielding an OCCC < 0.85) and multivariable analysis, with most features (1042/1414) revealing a systematic trend with the blending level (multiple comparisons adjusted p < 0.05). Reproducibility increased in association to image processing with smooth filters, nonetheless specific investigation in relation to clinical endpoints should be performed to ensure that textural information is not removed.

CONCLUSIONS

Combining univariate and multivariable models is allowed to identify features for which corrections may be applied to reduce the trend with the algorithm and increase reproducibility. Subsequent clustering may be applied to eliminate residual redundancy.

18F-FDOPA PET for the Noninvasive Prediction of Glioma Molecular Parameters: A Radiomics Study

The assessment of gliomas by 18F-FDOPA PET imaging as an adjunct to MRI showed high performance by combining static and dynamic features to noninvasively predict the isocitrate dehydrogenase (IDH) mutations and the 1p/19q codeletion, which the World Health Organization classified as significant parameters in 2016. The current study evaluated whether other 18F-FDOPA PET radiomics features further improve performance and the contributions of each of these features to performance. Methods: Our study included 72 retrospectively selected, newly diagnosed glioma patients with 18F-FDOPA PET dynamic acquisitions. A set of 114 features, including conventional static features and dynamic features, as well as other radiomics features, were extracted and machine-learning models trained to predict IDH mutations and the 1p/19q codeletion. Models were based on a machine-learning algorithm built from stable, relevant, and uncorrelated features selected by hierarchic clustering followed by a bootstrapped feature selection process. Models were assessed by comparing area under the curve using a nested cross-validation approach. Feature importance was assessed using Shapley additive explanations values. Results: The best models were able to predict IDH mutations (logistic regression with L2 regularization) and the 1p/19q codeletion (support vector machine with radial basis function kernel) with an area under the curve of 0.831 (95% CI, 0.790-0.873) and 0.724 (95% CI, 0.669-0.782), respectively. For the prediction of IDH mutations, dynamic features were the most important features in the model (time to peak, 35.5%). In contrast, other radiomics features were the most useful for predicting the 1p/19q codeletion (up to 14.5% of importance for the small-zone low-gray-level emphasis). Conclusion:18F-FDOPA PET is an effective tool for the noninvasive prediction of glioma molecular parameters using a full set of amino-acid PET radiomics features. The contribution of each feature set shows the importance of systematically integrating dynamic acquisition for prediction of the IDH mutations as well as developing the use of radiomics features in routine practice for prediction of the 1p/19q codeletion.

Effects of tracer uptake time in non-small cell lung cancer 18F-FDG PET radiomics

Positron emission tomography (PET) radiomics applied to oncology allows the measurement of intra-tumoral heterogeneity. This quantification can be affected by image protocols hence there is an increased interest in understanding how radiomic expression on PET images is affected by different imaging conditions. To address that, this study explores how radiomic features are affected by changes in ¹⁸F-FDG uptake time, image reconstruction, lesion delineation, and radiomics binning settings. Methods: Ten non-small cell lung cancer (NSCLC) patients underwent ¹⁸F-FDG PET scans on two consecutive days. On each day, scans were obtained at 60min and 90min post-injection and reconstructed following EARL version 1 (EARL1) and with point-spread-function resolution modelling (PSF-EARL2). Lesions were delineated using thresholds at SUV=4.0, 40% of SUV_max, and with a contrast-based isocontour. PET image intensity was discretized with both fixed bin width (FBW) and fixed bin number (FBN) before the calculation of the radiomic features. Repeatability of features was measured with intraclass correlation (ICC), and the change in feature value over time was calculated as a function of its repeatability. Features were then classified on use-case scenarios based on their repeatability and susceptibility to tracer uptake time. Results: With PSF-EARL2 reconstruction, 40% of SUV_max lesion delineation, and FBW intensity discretization, most features (94%) were repeatable at both uptake times (ICC>0.9), 39% being classified for dual-time-point use-case for being sensitive to changes in uptake time, 39% were classified for cross-sectional studies with unclear dependency on time, 20% classified for cross-sectional use while being robust to tracer uptake time changes, and 6% were discarded for poor repeatability. EARL1 images had one less repeatable feature than PSF-EARL2 (Neighborhood Gray-Level Different Matrix Coarseness), the contrast-based delineation had the poorest repeatability of the delineation methods with 45% features being discarded, and FBN resulted in lower repeatability than FBW (45% and 6% features were discarded, respectively). Conclusion: Repeatability was maximized with PSF-EARL2 reconstruction, lesion delineation at 40% of SUV_max, and FBW intensity discretization. Based on their susceptibility to tracer uptake time, radiomic features were classified into specific NSCLC PET radiomics use-cases.

Relevance of Dynamic 18F-DOPA PET Radiomics for Differentiation of High-Grade Glioma Progression from Treatment-Related Changes

This study evaluates the relevance of ¹⁸F-DOPA PET static and dynamic radiomics for differentiation of high-grade glioma (HGG) progression from treatment-related changes (TRC) by comparing diagnostic performances to the current PET imaging standard of care. Eighty-five patients with histologically confirmed HGG and investigated by dynamic ¹⁸F-FDOPA PET in two institutions were retrospectively selected. ElasticNet logistic regression, Random Forest and XGBoost machine models were trained with different sets of features-radiomics extracted from static tumor-to-background-ratio (TBR) parametric images, radiomics extracted from time-to-peak (TTP) parametric images, as well as combination of both-in order to discriminate glioma progression from TRC at 6 months from the PET scan. Diagnostic performances of the models were compared to a logistic regression model with TBR_mean ± clinical features used as reference. Training was performed on data from the first center, while external validation was performed on data from the second center. Best radiomics models showed only slightly better performances than the reference model (respective AUCs of 0.834 vs. 0.792, p < 0.001). Our current results show similar findings at the multicentric level using different machine learning models and report a marginal additional value for TBR static and TTP dynamic radiomics over the classical analysis based on TBR values.

Evaluation of conventional and deep learning based image harmonization methods in radiomics studies

Objective.To evaluate the impact of image harmonization on outcome prediction models using radiomics.Approach.234 patients from the Brain Tumor Image Segmentation Benchmark (BRATS) dataset with T1 MRI were enrolled in this study. Images were harmonized to a reference image using histogram matching (H_HM) and a generative adversarial network (GAN)-based method (H_GAN). 88 radiomics features were extracted on H_HM, H_GANand original (H_NONE) images. Wilcoxon paired test was used to identify features significantly impacted by the harmonization protocol used. Radiomic prediction models were built using feature selection with the Least Absolute Shrinkage and Selection Operator (LASSO) and Kaplan-Meier analysis.Main results.More than 50% of the features (49/88) were statistically modified by the harmonization with H_HMand 55 with H_GAN(adjustedp-value < 0.05). The contribution of histogram and texture features selected by the LASSO, in comparison to shape features that were not impacted by harmonization, was higher in harmonized datasets (47% for H_none, 62% for H_HMand 71% for H_GAN). Both image-based harmonization methods allowed to split patients into two groups with significantly different survival (p<0.05). With the H_GANimages, we were also able to build and validate a model using only features impacted by the harmonization (median survivals of 189 versus 437 days,p= 0.006)Significance.Data harmonization in a multi-institutional cohort allows to recover the predictive value of some radiomics features that was lost due to differences in the image properties across centers. In terms of ability to build survival prediction models in the BRATS dataset, the loss of power from impacted histogram and heterogeneity features was compensated by the selection of additional shape features. The harmonization using a GAN-based approach outperformed the histogram matching technique, supporting the interest for the development of new advanced harmonization techniques for radiomic analysis purposes.

Convolutional Neural Network Addresses the Confounding Impact of CT Reconstruction Kernels on Radiomics Studies

Achieving high feature reproducibility while preserving biological information is one of the main challenges for the generalizability of current radiomics studies. Non-clinical imaging variables, such as reconstruction kernels, have shown to significantly impact radiomics features. In this study, we retrain an open-source convolutional neural network (CNN) to harmonize computerized tomography (CT) images with various reconstruction kernels to improve feature reproducibility and radiomic model performance using epidermal growth factor receptor (EGFR) mutation prediction in lung cancer as a paradigm. In the training phase, the CNN was retrained and tested on 32 lung cancer patients’ CT images between two different groups of reconstruction kernels (smooth and sharp). In the validation phase, the retrained CNN was validated on an external cohort of 223 lung cancer patients’ CT images acquired using different CT scanners and kernels. The results showed that the retrained CNN could be successfully applied to external datasets with different CT scanner parameters, and harmonization of reconstruction kernels from sharp to smooth could significantly improve the performance of radiomics model in predicting EGFR mutation status in lung cancer. In conclusion, the CNN based method showed great potential in improving feature reproducibility and generalizability by harmonizing medical images with heterogeneous reconstruction kernels.

18F-FDG texture analysis predicts the pathological Fuhrman nuclear grade of clear cell renal cell carcinoma

PURPOSE

This article analyzes the image heterogeneity of clear cell renal cell carcinoma (ccRCC) based on positron emission tomography (PET) and positron emission tomography-computed tomography (PET/CT) texture parameters, and provides a new objective quantitative parameter for predicting pathological Fuhrman nuclear grading before surgery.

METHODS

A retrospective analysis was performed on preoperative PET/CT images of 49 patients whose surgical pathology was ccRCC, 27 of whom were low grade (Fuhrman I/II) and 22 of whom were high grade (Fuhrman III/IV). Radiological parameters and standard uptake value (SUV) indicators on PET and computed tomography (CT) images were extracted by using the LIFEx software package. The discriminative ability of each texture parameter was evaluated through receiver operating curve (ROC). Binary logistic regression analysis was used to screen the texture parameters with distinguishing and diagnostic capabilities and whose area under curve (AUC) > 0.5. DeLong's test was used to compare the AUCs of PET texture parameter model and PET/CT texture parameter model with traditional maximum standardized uptake value (SUVmax) model and the ratio of tumor SUVmax to liver SUVmean (SUL)model. In addition, the models with the larger AUCs among the SUV models and texture models were prospectively internally verified.

RESULTS

In the ROC curve analysis, the AUCs of SUVmax model, SUL model, PET texture parameter model, and PET/CT texture parameter model were 0.803, 0.819, 0.873, and 0.926, respectively. The prediction ability of PET texture parameter model or PET/CT texture parameter model was significantly better than SUVmax model (P = 0.017, P = 0.02), but it was not better than SUL model (P = 0.269, P = 0.053). In the prospective validation cohort, both the SUL model and the PET/CT texture parameter model had good predictive ability, and the AUCs of them were 0.727 and 0.792, respectively.

CONCLUSION

PET and PET/CT texture parameter models can improve the prediction ability of ccRCC Fuhrman nuclear grade; SUL model may be the more accurate and easiest way to predict ccRCC Fuhrman nuclear grade.

A systematic review and quality of reporting checklist for repeatability and reproducibility of radiomic features

•

Main factors impacting feature stability: Image acquisition, reconstruction, tumor segmentation, and interpolation.

•

Textural features are less robust than morphological or statistical features.

•

A checklist is provided including items that should be reported in a radiomic study.

Purpose

Although quantitative image biomarkers (radiomics) show promising value for cancer diagnosis, prognosis, and treatment assessment, these biomarkers still lack reproducibility. In this systematic review, we aimed to assess the progress in radiomics reproducibility and repeatability in the recent years.

Methods and materials

Four hundred fifty-one abstracts were retrieved according to the original PubMed search pattern with the publication dates ranging from 2017/05/01 to 2020/12/01. Each abstract including the keywords was independently screened by four observers. Forty-two full-text articles were selected for further analysis. Patient population data, radiomic feature classes, feature extraction software, image preprocessing, and reproducibility results were extracted from each article. To support the community with a standardized reporting strategy, we propose a specific reporting checklist to evaluate the feasibility to reproduce each study.

Results

Many studies continue to under-report essential reproducibility information: all but one clinical and all but two phantom studies missed to report at least one important item reporting image acquisition. The studies included in this review indicate that all radiomic features are sensitive to image acquisition, reconstruction, tumor segmentation, and interpolation. However, the amount of sensitivity is feature dependent, for instance, textural features were, in general, less robust than statistical features.

Conclusions

Radiomics repeatability, reproducibility, and reporting quality can substantially be improved regarding feature extraction software and settings, image preprocessing and acquisition, cutoff values for stable feature selection. Our proposed radiomics reporting checklist can serve to simplify and improve the reporting and, eventually, guarantee the possibility to fully replicate and validate radiomic studies.

CT-based radiomics stratification of tumor grade and TNM stage of clear cell renal cell carcinoma

OBJECTIVES

To evaluate the utility of CT-based radiomics signatures in discriminating low-grade (grades 1-2) clear cell renal cell carcinomas (ccRCC) from high-grade (grades 3-4) and low TNM stage (stages I-II) ccRCC from high TNM stage (stages III-IV).

METHODS

A total of 587 subjects (mean age 60.2 years ± 12.2; range 22-88.7 years) with ccRCC were included. A total of 255 tumors were high grade and 153 were high stage. For each subject, one dominant tumor was delineated as the region of interest (ROI). Our institutional radiomics pipeline was then used to extract 2824 radiomics features across 12 texture families from the manually segmented volumes of interest. Separate iterations of the machine learning models using all extracted features (full model) as well as only a subset of previously identified robust metrics (robust model) were developed. Variable of importance (VOI) analysis was performed using the out-of-bag Gini index to identify the top 10 radiomics metrics driving each classifier. Model performance was reported using area under the receiver operating curve (AUC).

RESULTS

The highest AUC to distinguish between low- and high-grade ccRCC was 0.70 (95% CI 0.62-0.78) and the highest AUC to distinguish between low- and high-stage ccRCC was 0.80 (95% CI 0.74-0.86). Comparable AUCs of 0.73 (95% CI 0.65-0.8) and 0.77 (95% CI 0.7-0.84) were reported using the robust model for grade and stage classification, respectively. VOI analysis revealed the importance of neighborhood operation-based methods, including GLCM, GLDM, and GLRLM, in driving the performance of the robust models for both grade and stage classification.

CONCLUSION

Post-validation, CT-based radiomics signatures may prove to be useful tools to assess ccRCC grade and stage and could potentially add to current prognostic models. Multiphase CT-based radiomics signatures have potential to serve as a non-invasive stratification schema for distinguishing between low- and high-grade as well as low- and high-stage ccRCC.

KEY POINTS

• Radiomics signatures derived from clinical multiphase CT images were able to stratify low- from high-grade ccRCC, with an AUC of 0.70 (95% CI 0.62-0.78). • Radiomics signatures derived from multiphase CT images yielded discriminative power to stratify low from high TNM stage in ccRCC, with an AUC of 0.80 (95% CI 0.74-0.86). • Models created using only robust radiomics features achieved comparable AUCs of 0.73 (95% CI 0.65-0.80) and 0.77 (95% CI 0.70-0.84) to the model with all radiomics features in classifying ccRCC grade and stage, respectively.

Radiomics for Identification and Prediction in Metastatic Prostate Cancer: A Review of Studies

Metastatic Prostate Cancer (mPCa) is associated with a poor patient prognosis. mPCa spreads throughout the body, often to bones, with spatial and temporal variations that make the clinical management of the disease difficult. The evolution of the disease leads to spatial heterogeneity that is extremely difficult to characterise with solid biopsies. Imaging provides the opportunity to quantify disease spread. Advanced image analytics methods, including radiomics, offer the opportunity to characterise heterogeneity beyond what can be achieved with simple assessment. Radiomics analysis has the potential to yield useful quantitative imaging biomarkers that can improve the early detection of mPCa, predict disease progression, assess response, and potentially inform the choice of treatment procedures. Traditional radiomics analysis involves modelling with hand-crafted features designed using significant domain knowledge. On the other hand, artificial intelligence techniques such as deep learning can facilitate end-to-end automated feature extraction and model generation with minimal human intervention. Radiomics models have the potential to become vital pieces in the oncology workflow, however, the current limitations of the field, such as limited reproducibility, are impeding their translation into clinical practice. This review provides an overview of the radiomics methodology, detailing critical aspects affecting the reproducibility of features, and providing examples of how artificial intelligence techniques can be incorporated into the workflow. The current landscape of publications utilising radiomics methods in the assessment and treatment of mPCa are surveyed and reviewed. Associated studies have incorporated information from multiple imaging modalities, including bone scintigraphy, CT, PET with varying tracers, multiparametric MRI together with clinical covariates, spanning the prediction of progression through to overall survival in varying cohorts. The methodological quality of each study is quantified using the radiomics quality score. Multiple deficits were identified, with the lack of prospective design and external validation highlighted as major impediments to clinical translation. These results inform some recommendations for future directions of the field.

Radiomics integration into a picture archiving and communication system

Radiomics is referred to as quantitative imaging of biomarkers used for clinical outcome prognosis or tumor characterization. In order to bridge radiomics and its clinical application, we aimed to build an integrated solution of radiomics extraction with an open-source Picture Archiving and Communication System (PACS). The integrated SQLite4Radiomics software was tested in three different imaging modalities and its performance was benchmarked in lung cancer open datasets RIDER and MMD with median extraction time of 10.7 (percentiles 25-75: 8.9-18.7) seconds per ROI in three different configurations.

Predicting clinical outcomes in COVID-19 using radiomics on chest radiographs

Objectives

For optimal utilization of healthcare resources, there is a critical need for early identification of COVID-19 patients at risk of poor prognosis as defined by the need for intensive unit care and mechanical ventilation. We tested the feasibility of chest X-ray (CXR)-based radiomics metrics to develop machine-learning algorithms for predicting patients with poor outcomes.

Methods

In this Institutional Review Board (IRB) approved, Health Insurance Portability and Accountability Act (HIPAA) compliant, retrospective study, we evaluated CXRs performed around the time of admission from 167 COVID-19 patients. Of the 167 patients, 68 (40.72%) required intensive care during their stay, 45 (26.95%) required intubation, and 25 (14.97%) died. Lung opacities were manually segmented using ITK-SNAP (open-source software). CaPTk (open-source software) was used to perform 2D radiomics analysis.

Results

Of all the algorithms considered, the AdaBoost classifier performed the best with AUC = 0.72 to predict the need for intubation, AUC = 0.71 to predict death, and AUC = 0.61 to predict the need for admission to the intensive care unit (ICU). AdaBoost had similar performance with ElasticNet in predicting the need for admission to ICU. Analysis of the key radiomic metrics that drive model prediction and performance showed the importance of first-order texture metrics compared to other radiomics panel metrics. Using a Venn-diagram analysis, two first-order texture metrics and one second-order texture metric that consistently played an important role in driving model performance in all three outcome predictions were identified.

Conclusions:

Considering the quantitative nature and reliability of radiomic metrics, they can be used prospectively as prognostic markers to individualize treatment plans for COVID-19 patients and also assist with healthcare resource management.

Advances in knowledge

We report on the performance of CXR-based imaging metrics extracted from RT-PCR positive COVID-19 patients at admission to develop machine-learning algorithms for predicting the need for ICU, the need for intubation, and mortality, respectively.

[18F]FDG PET radiomics to predict disease-free survival in cervical cancer: a multi-scanner/center study with external validation

PURPOSE

To test the performances of native and tumour to liver ratio (TLR) radiomic features extracted from pre-treatment 2-[18F] fluoro-2-deoxy-D-glucose ([18F]FDG) PET/CT and combined with machine learning (ML) for predicting cancer recurrence in patients with locally advanced cervical cancer (LACC).

METHODS

One hundred fifty-eight patients with LACC from multiple centers were retrospectively included in the study. Tumours were segmented using the Fuzzy Local Adaptive Bayesian (FLAB) algorithm. Radiomic features were extracted from the tumours and from regions drawn over the normal liver. Cox proportional hazard model was used to test statistical significance of clinical and radiomic features. Fivefold cross validation was used to tune the number of features. Seven different feature selection methods and four classifiers were tested. The models with the selected features were trained using bootstrapping and tested in data from each scanner independently. Reproducibility of radiomics features, clinical data added value and effect of ComBat-based harmonisation were evaluated across scanners.

RESULTS

After a median follow-up of 23 months, 29% of the patients recurred. No individual radiomic or clinical features were significantly associated with cancer recurrence. The best model was obtained using 10 TLR features combined with clinical information. The area under the curve (AUC), F1-score, precision and recall were respectively 0.78 (0.67-0.88), 0.49 (0.25-0.67), 0.42 (0.25-0.60) and 0.63 (0.20-0.80). ComBat did not improve the predictive performance of the best models. Both the TLR and the native models performance varied across scanners used in the test set.

CONCLUSION

[18F]FDG PET radiomic features combined with ML add relevant information to the standard clinical parameters in terms of LACC patient's outcome but remain subject to variability across PET/CT devices.

Radiomics in PET Imaging:: A Practical Guide for Newcomers

Radiomics has undergone considerable development in recent years. In PET imaging, very promising results concerning the ability of handcrafted features to predict the biological characteristics of lesions and to assess patient prognosis or response to treatment have been reported in the literature. This article presents a checklist for designing a reliable radiomic study, gives an overview of the steps of the pipeline, and outlines approaches for data harmonization. Tips are provided for critical reading of the content of articles. The advantages and limitations of handcrafted radiomics compared with deep-learning approaches for the characterization of PET images are also discussed.

Multi-Stage Harmonization for Robust AI across Breast MR Databases

Simple Summary

Batch harmonization of radiomic features extracted from magnetic resonance images of breast lesions from two databases was applied to an artificial intelligence/machine learning classification workflow. Training and independent test sets from the two databases, as well as the combination of them, were used in pre-harmonization and post-harmonization forms to investigate the generalizability of performance in the task of distinguishing between malignant and benign lesions. Most training and independent test scenarios were statistically equivalent, demonstrating that batch harmonization with feature selection harmonization can potentially develop generalizable classification models.

Abstract

Radiomic features extracted from medical images may demonstrate a batch effect when cases come from different sources. We investigated classification performance using training and independent test sets drawn from two sources using both pre-harmonization and post-harmonization features. In this retrospective study, a database of thirty-two radiomic features, extracted from DCE-MR images of breast lesions after fuzzy c-means segmentation, was collected. There were 944 unique lesions in Database A (208 benign lesions, 736 cancers) and 1986 unique lesions in Database B (481 benign lesions, 1505 cancers). The lesions from each database were divided by year of image acquisition into training and independent test sets, separately by database and in combination. ComBat batch harmonization was conducted on the combined training set to minimize the batch effect on eligible features by database. The empirical Bayes estimates from the feature harmonization were applied to the eligible features of the combined independent test set. The training sets (A, B, and combined) were then used in training linear discriminant analysis classifiers after stepwise feature selection. The classifiers were then run on the A, B, and combined independent test sets. Classification performance was compared using pre-harmonization features to post-harmonization features, including their corresponding feature selection, evaluated using the area under the receiver operating characteristic curve (AUC) as the figure of merit. Four out of five training and independent test scenarios demonstrated statistically equivalent classification performance when compared pre- and post-harmonization. These results demonstrate that translation of machine learning techniques with batch data harmonization can potentially yield generalizable models that maintain classification performance.

FDG PET biomarkers for prediction of survival in metastatic melanoma prior to anti-PD1 immunotherapy

Our aim was to analyse whether biomarkers extracted from baseline ¹⁸F-FDG PET before anti-PD1 treatment contribute to prognostic survival information for early risk stratification in metastatic melanoma. Fifty-six patients, without prior systemic treatment, BRAF wild type, explored using ¹⁸F-FDG PET were included retrospectively. Our primary endpoint was overall survival (OS). Total metabolic tumoral volume (MTV) and forty-one IBSI compliant parameters were extracted from PET. Parameters associated with outcome were evaluated by a cox regression model and when significant helped build a prognostic score. Median follow-up was 22.1 months and 21 patients died. Total MTV and long zone emphasis (LZE) correlated with shorter OS and served to define three risk categories for the prognostic score. For low, intermediate and high risk groups, survival rates were respectively 91.1% (IC 95 80–1), 56.1% (IC 95 37.1–85) and 19% (IC 95 0.06–60.2) and hazard ratios were respectively 0.11 (IC 95 0.025–0.46), P = 0.0028, 1.2 (IC 95 0.48–2.8), P = 0.74 and 5.9 (IC 95 2.5–14), P < 0.0001. To conclude, a prognostic score based on total MTV and LZE separated metastatic melanoma patients in 3 categories with dramatically different outcomes. Innovative therapies should be tested in the group with the lowest prognosis score for future clinical trials.

A guide to ComBat harmonization of imaging biomarkers in multicenter studies

The impact of PET image acquisition and reconstruction parameters on SUV measurements or radiomic feature values is widely documented. This "scanner" effect is detrimental to the design and validation of predictive or prognostic models and limits the use of large multicenter cohorts. To reduce the impact of this scanner effect, the ComBat method has been proposed and is now used in various contexts. The purpose of this article is to explain and illustrate the use of ComBat based on practical examples. We also give examples in which the ComBat assumptions are not met; thus, ComBat should not be used.

Prediction of TERTp-mutation status in IDH-wildtype high-grade gliomas using pre-treatment dynamic [18F]FET PET radiomics

Purpose

To evaluate radiomic features extracted from standard static images (20–40 min p.i.), early summation images (5–15 min p.i.), and dynamic [¹⁸F]FET PET images for the prediction of TERTp-mutation status in patients with IDH-wildtype high-grade glioma.

Methods

A total of 159 patients (median age 60.2 years, range 19–82 years) with newly diagnosed IDH-wildtype diffuse astrocytic glioma (WHO grade III or IV) and dynamic [¹⁸F]FET PET prior to surgical intervention were enrolled and divided into a training (n = 112) and a testing cohort (n = 47) randomly. First-order, shape, and texture radiomic features were extracted from standard static (20–40 min summation images; TBR_20–40), early static (5–15 min summation images; TBR_5–15), and dynamic (time-to-peak; TTP) images, respectively. Recursive feature elimination was used for feature selection by 10-fold cross-validation in the training cohort after normalization, and logistic regression models were generated using the radiomic features extracted from each image to differentiate TERTp-mutation status. The areas under the ROC curve (AUC), accuracy, sensitivity, specificity, and positive and negative predictive value were calculated to illustrate diagnostic power in both the training and testing cohort.

Results

The TTP model comprised nine selected features and achieved highest predictability of TERTp-mutation with an AUC of 0.82 (95% confidence interval 0.71–0.92) and sensitivity of 92.1% in the independent testing cohort. Weak predictive capability was obtained in the TBR_5–15 model, with an AUC of 0.61 (95% CI 0.42–0.80) in the testing cohort, while no predictive power was observed in the TBR_20–40 model.

Conclusions

Radiomics based on TTP images extracted from dynamic [¹⁸F]FET PET can predict the TERTp-mutation status of IDH-wildtype diffuse astrocytic high-grade gliomas with high accuracy preoperatively.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00259-021-05526-6.

Multi-parametric MRI phenotype with trustworthy machine learning for differentiating CNS demyelinating diseases

Background

Misdiagnosis of multiple sclerosis (MS) and neuromyelitis optica (NMO) may delay the treatment, resulting in poor prognosis. However, the precise identification of these two diseases is still challenging in clinical practice. We aimed to evaluate the value of quantitative radiomic features extracted from the brain white matter lesions for differential diagnosis of MS and NMO.

Methods

We recruited 116 CNS demyelinating patients including 78 MS, and 38 NMO. Three neuroradiologists performed visual differential diagnosis based on brain MRI for comparison purpose. A multi-level scheme was designed to harness the selection of discriminative and stable radiomics features extracted from brain while mater lesions in T1-MPRAGE, T2 sequences and clinical factors. Based on the imaging phenotype composed of the selected radiomic and clinical features, Multi-parametric Multivariate Random Forest (MM-RF) model was constructed and verified with both 10-fold cross-validation and independent testing. Result interpretation was provided to build trust in diagnostic decisions.

Results

Eighty-six patients were randomly selected to form the training set while the rest 30 patients for independent testing. On the training set, our MM-RF model achieved accuracy 0.849 and AUC 0.826 in 10-fold cross-validation, which were significantly higher than clinical visual analysis (0.709 and 0.683, p < 0.05). In the independent testing, the MM-RF model achieved AUC 0.902, accuracy 0.871, sensitivity 0.873, specificity 0.869, respectively. Furthermore, age, sex and EDSS were found mildly correlated with the radiomic features (p of all < 0.05).

Conclusions

Multi-parametric radiomic features have potential as practical quantitative imaging biomarkers for differentiating MS from NMO.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-021-03015-w.

Making Radiomics More Reproducible across Scanner and Imaging Protocol Variations: A Review of Harmonization Methods

Radiomics converts medical images into mineable data via a high-throughput extraction of quantitative features used for clinical decision support. However, these radiomic features are susceptible to variation across scanners, acquisition protocols, and reconstruction settings. Various investigations have assessed the reproducibility and validation of radiomic features across these discrepancies. In this narrative review, we combine systematic keyword searches with prior domain knowledge to discuss various harmonization solutions to make the radiomic features more reproducible across various scanners and protocol settings. Different harmonization solutions are discussed and divided into two main categories: image domain and feature domain. The image domain category comprises methods such as the standardization of image acquisition, post-processing of raw sensor-level image data, data augmentation techniques, and style transfer. The feature domain category consists of methods such as the identification of reproducible features and normalization techniques such as statistical normalization, intensity harmonization, ComBat and its derivatives, and normalization using deep learning. We also reflect upon the importance of deep learning solutions for addressing variability across multi-centric radiomic studies especially using generative adversarial networks (GANs), neural style transfer (NST) techniques, or a combination of both. We cover a broader range of methods especially GANs and NST methods in more detail than previous reviews.

A review in radiomics: Making personalized medicine a reality via routine imaging

Radiomics is the quantitative analysis of standard-of-care medical imaging; the information obtained can be applied within clinical decision support systems to create diagnostic, prognostic, and/or predictive models. Radiomics analysis can be performed by extracting hand-crafted radiomics features or via deep learning algorithms. Radiomics has evolved tremendously in the last decade, becoming a bridge between imaging and precision medicine. Radiomics exploits sophisticated image analysis tools coupled with statistical elaboration to extract the wealth of information hidden inside medical images, such as computed tomography (CT), magnetic resonance (MR), and/or Positron emission tomography (PET) scans, routinely performed in the everyday clinical practice. Many efforts have been devoted in recent years to the standardization and validation of radiomics approaches, to demonstrate their usefulness and robustness beyond any reasonable doubts. However, the booming of publications and commercial applications of radiomics approaches warrant caution and proper understanding of all the factors involved to avoid "scientific pollution" and overly enthusiastic claims by researchers and clinicians alike. For these reasons the present review aims to be a guidebook of sorts, describing the process of radiomics, its pitfalls, challenges, and opportunities, along with its ability to improve clinical decision-making, from oncology and respiratory medicine to pharmacological and genotyping studies.

Radiomics in the Setting of Neoadjuvant Radiotherapy: A New Approach for Tailored Treatment

INTRODUCTION

Neoadjuvant radiotherapy is currently used mainly in locally advanced rectal cancer and sarcoma and in a subset of non-small cell lung cancer and esophageal cancer, whereas in other diseases it is under investigation. The evaluation of the efficacy of the induction strategy is made possible by performing imaging investigations before and after the neoadjuvant therapy and is usually challenging. In the last decade, texture analysis (TA) has been developed to help the radiologist to quantify and identify the parameters related to tumor heterogeneity, which cannot be appreciated by the naked eye. The aim of this narrative is to review the impact of TA on the prediction of response to neoadjuvant radiotherapy and or chemoradiotherapy.

MATERIALS AND METHODS

Key references were derived from a PubMed query. Hand searching and ClinicalTrials.gov were also used.

RESULTS

This paper contains a narrative report and a critical discussion of radiomics approaches in different fields of neoadjuvant radiotherapy, including esophageal cancer, lung cancer, sarcoma, and rectal cancer.

CONCLUSIONS

Radiomics can shed a light on the setting of neoadjuvant therapies that can be used to tailor subsequent approaches or even to avoid surgery in the future. At the same, these results need to be validated in prospective and multicenter trials.

Quantitative radiomics features in diffuse large B-cell lymphoma: does segmentation method matter?

INTRODUCTION

Radiomics features may predict outcome in diffuse large B-cell lymphoma (DLBCL). Currently, multiple segmentation methods are used to calculate metabolic tumor volume (MTV). We assessed the influence of segmentation method on the discriminative power of radiomics features in DLBCL for patient level and for the largest lesion. Methods: 50 baseline ¹⁸F-fluorodeoxyglucose positron emission tomography computed tomography (PET/CT) scans of DLBCL patients who progressed or relapsed within 2 years after diagnosis were matched on uptake time and reconstruction method with 50 baseline PET/CT scans of DLBCL patients without progression. Scans were analysed using 6 semi-automatic segmentation methods (standardized uptake value (SUV)4.0, SUV2.5, 41% of the maximum SUV, 50% of the SUVpeak, majority vote (MV)2 and MV3, respectively). Based on these segmentations, 490 radiomics features were extracted at patient level and 486 features for the largest lesion. To quantify the agreement between features extracted from different segmentation methods, the intra-class correlation (ICC) agreement was calculated for each method compared to SUV4.0. The feature space was reduced by deleting features that had high Pearson correlations (≥0.7) with the previously established predictors MTV and/or SUVpeak. Model performance was assessed using stratified repeated cross-validation with 5 folds and 2000 repeats yielding the mean receiver-operating characteristics curve integral (CV-AUC) for all segmentation methods using logistic regression with backward feature selection. Results: The percentage of features yielding an ICC ≥0.75 compared to the SUV4.0 segmentation was lowest for A50P both at patient level and for the largest lesion, with 77.3% and 66.7% of the features yielding an ICC ≥0.75, respectively. Features were not highly correlated with MTV, with at least 435 features at patient level and 409 features for the largest lesion for all segmentation methods with a correlation coefficient <0.7. Features were highly correlated with SUVpeak (at least 190 and 134 were uncorrelated, respectively). CV-AUCs ranged between 0.69±0.11 and 0.84±0.09 for patient level, and between 0.69±0.11 and 0.73±0.10 for lesion level. Conclusion: Even though there are differences in the actual radiomics feature values derived and selected features between segmentation methods, there is no substantial difference in the discriminative power of radiomics features between segmentation methods.

Radiomic assessment as a method for predicting tumor mutation burden (TMB) of bladder cancer patients: a feasibility study

Background

Tumor mutation burden (TMB) is an emerging prognostic biomarker of immunotherapy for bladder cancer (BLCA). We aim at investigating radiomic features’ value in predicting the TMB status of BLCA patients.

Methods

Totally, 75 patients with BLCA were enrolled. Radiomic features extracted from the volume of interest of preoperative pelvic contrast-enhanced computed tomography (CECT) were obtained for each case. Unsupervised hierarchical clustering analysis was performed based on radiomic features. Sequential univariate Logistic regression, the least absolute shrinkage and selection operator (LASSO) regression and the backward stepwise regression were used to develop a TMB-predicting model using radiomic features.

Results

The unsupervised clustering analysis divided the total cohort into two groups, i.e., group A (32.0%) and B (68.0%). Patients in group A had a significantly larger proportion of having high TMB against those in group B (66.7% vs. 41.2%, p = 0.039), indicating the intrinsic ability of radiomic features in TMB-predicting. In univariate analysis, 27 radiomic features could predict TMB. Based on six radiomic features selected by logistic and LASSO regression, a TMB-predicting model was built and visualized by nomogram. The area under the ROC curve of the model reached 0.853. Besides, the calibration curve and the decision curve also revealed the good performance of the model.

Conclusions

Our work firstly proved the feasibility of using radiomics to predict TMB for patients with BLCA. The predictive model based on radiomic features from pelvic CECT has a promising ability to predict TMB. Future study with a larger cohort is needed to verify our findings.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-021-08569-y.

Statistical harmonization can improve the development of a multicenter CT-based radiomic model predictive of nonresponse to induction chemotherapy in laryngeal cancers

PURPOSE

To develop a radiomic model predicting nonresponse to induction chemotherapy in laryngeal cancers, from multicenter pretherapeutic contrast-enhanced computed tomography (CE-CT) and evaluate the benefit of feature harmonization in such a context.

METHODS

Patients (n = 104) eligible for laryngeal preservation chemotherapy were included in five centers. Primary tumor was manually delineated on the CE-CT images. The following radiomic features were extracted with an in-house software (MIRAS v1.1, LaTIM UMR 1101): intensity, shape, and textural features derived from Gray-Level Co-occurrence Matrix (GLCM), Neighborhood Gray Tone Difference Matrix (NGTDM), Gray-Level Run Length Matrix (GLRLM), and Gray-Level Size Zone Matrix (GLSZM). Harmonization was performed using ComBat after unsupervised hierarchical clustering, used to determine labels automatically, given the high heterogeneity of imaging characteristics across and within centers. Patients with similar feature distributions were grouped with unsupervised clustering into an optimal number of clusters (2) determined with "silhouette scoring." Statistical harmonization was then carried out with ComBat on these 2 identified clusters. The cohort was split into training/validation (n = 66) and testing (n = 32) sets. Area under the receiver operating characteristics curves (AUC) were used to evaluate the ability of radiomic features (before and after harmonization) to predict nonresponse to chemotherapy, and specificity (Sp) and sensitivity (Se) were used to quantify their performance in the testing set.

RESULTS

Without harmonization, none of the features identified as predictive in the training set remained significant in the testing set. After ComBat, one textural feature identified in the training set keeps a predictive trend in the testing set-Zone Percentage, derived from the GLSZM, was predictive of nonresponse in the training set (AUC = 0.62, Se = 70%, Sp = 64%, P = 0.04) and obtained a satisfactory performance in the testing set (Se = 80%, Sp = 67%, P = 0.03), although significance was limited by the size of the testing set. These results are consistent with previously published findings in head and neck cancers.

CONCLUSIONS

Radiomic features from CE-CT could help in the selection of patients for induction chemotherapy in laryngeal cancers, with relatively good sensitivity and specificity in predicting lack of response. Statistical harmonization with ComBat and unsupervised clustering seems to improve the predictive value of features extracted in such a heterogeneous multicenter setting.

Prognostic value of 18F-FDG PET/CT with texture analysis in patients with rectal cancer treated by surgery

PURPOSE

The aim of this study was to evaluate the ability of texture analysis using pretreatment 18F-FDG PET/CT to predict prognosis in patients with surgically treated rectal cancer.

METHODS

We analyzed 94 patients with pathologically proven rectal cancer who underwent pretreatment 18F-FDG PET/CT and were subsequently treated with surgery. The volume of interest of the primary tumor was defined using a threshold of 40% of the maximum standardized uptake value (SUVmax), and conventional (SUVmax, metabolic tumor volume [MTV], total lesion glycolysis [TLG]) and textural PET features were extracted. Harmonization of PET features was performed with the ComBat method. The study endpoints were overall survival (OS) and progression-free survival (PFS), and the prognostic value of PET features was evaluated by Cox regression analysis.

RESULTS

In the follow-up period (median 41.7 [interquartile range, 30.5-60.4] months), 21 (22.3%) and 30 (31.9%) patients had cancer-related death or disease progression, respectively. Univariate analysis revealed a significant association of (1) MTV, TLG, and gray-level co-occurrence matrix (GLCM) entropy with OS; and (2) SUVmax, MTV, TLG, and GLCM entropy with PFS. In multivariate analysis including clinical characteristics, GLCM entropy (≥ 2.13) was the only relevant prognostic PET feature for poor OS (hazard ratio [HR]: 4.16, p = 0.035) and PFS (HR: 2.70, p = 0.046).

CONCLUSION

GLCM entropy, which indicates metabolic intratumoral heterogeneity, was an independent prognostic factor in patients with surgically treated rectal cancer. Compared with conventional PET features, GLCM entropy has better predictive value and shows potential to facilitate precision medicine.

A transfer learning approach to facilitate ComBat-based harmonization of multicentre radiomic features in new datasets

PURPOSE

To facilitate the demonstration of the prognostic value of radiomics, multicenter radiomics studies are needed. Pooling radiomic features of such data in a statistical analysis is however challenging, as they are sensitive to the variability in scanner models, acquisition protocols and reconstruction settings, which is often unavoidable in a multicentre retrospective analysis. A statistical harmonization strategy called ComBat was utilized in radiomics studies to deal with the "center-effect". The goal of the present work was to integrate a transfer learning (TL) technique within ComBat-and recently developed alternate versions of ComBat with improved flexibility (M-ComBat) and robustness (B-ComBat)-to allow the use of a previously determined harmonization transform to the radiomic feature values of new patients from an already known center.

MATERIAL AND METHODS

The proposed TL approach were incorporated in the four versions of ComBat (standard, B, M, and B-M ComBat). The proposed approach was evaluated using a dataset of 189 locally advanced cervical cancer patients from 3 centers, with magnetic resonance imaging (MRI) and positron emission tomography (PET) images, with the clinical endpoint of predicting local failure. The impact performance of the TL approach was evaluated by comparing the harmonization achieved using only parts of the data to the reference (harmonization achieved using all the available data). It was performed through three different machine learning pipelines.

RESULTS

The proposed TL technique was successful in harmonizing features of new patients from a known center in all versions of ComBat, leading to predictive models reaching similar performance as the ones developed using the features harmonized with all the data available.

CONCLUSION

The proposed TL approach enables applying a previously determined ComBat transform to new, previously unseen data.

The Challenge of Choosing the Best Classification Method in Radiomic Analyses: Recommendations and Applications to Lung Cancer CT Images

Radiomics uses high-dimensional sets of imaging features to predict biological characteristics of tumors and clinical outcomes. The choice of the algorithm used to analyze radiomic features and perform predictions has a high impact on the results, thus the identification of adequate machine learning methods for radiomic applications is crucial. In this study we aim to identify suitable approaches of analysis for radiomic-based binary predictions, according to sample size, outcome balancing and the features-outcome association strength. Simulated data were obtained reproducing the correlation structure among 168 radiomic features extracted from Computed Tomography images of 270 Non-Small-Cell Lung Cancer (NSCLC) patients and the associated to lymph node status. Performances of six classifiers combined with six feature selection (FS) methods were assessed on the simulated data using AUC (Area Under the Receiver Operating Characteristics Curves), sensitivity, and specificity. For all the FS methods and regardless of the association strength, the tree-based classifiers Random Forest and Extreme Gradient Boosting obtained good performances (AUC ≥ 0.73), showing the best trade-off between sensitivity and specificity. On small samples, performances were generally lower than in large-medium samples and with larger variations. FS methods generally did not improve performances. Thus, in radiomic studies, we suggest evaluating the choice of FS and classifiers, considering specific sample size, balancing, and association strength.