Weakly Supervised MRI Slice-Level Deep Learning Classification of Prostate Cancer Approximates Full Voxel- and Slice-Level Annotation: Effect of Increasing Training Set Size

BACKGROUND

Weakly supervised learning promises reduced annotation effort while maintaining performance.

PURPOSE

To compare weakly supervised training with full slice-wise annotated training of a deep convolutional classification network (CNN) for prostate cancer (PC).

STUDY TYPE

Retrospective.

SUBJECTS

One thousand four hundred eighty-nine consecutive institutional prostate MRI examinations from men with suspicion for PC (65 ± 8 years) between January 2015 and November 2020 were split into training (N = 794, enriched with 204 PROSTATEx examinations) and test set (N = 695).

FIELD STRENGTH/SEQUENCE

1.5 and 3T, T2-weighted turbo-spin-echo and diffusion-weighted echo-planar imaging.

ASSESSMENT

Histopathological ground truth was provided by targeted and extended systematic biopsy. Reference training was performed using slice-level annotation (SLA) and compared to iterative training utilizing patient-level annotations (PLAs) with supervised feedback of CNN estimates into the next training iteration at three incremental training set sizes (N = 200, 500, 998). Model performance was assessed by comparing specificity at fixed sensitivity of 0.97 [254/262] emulating PI-RADS ≥ 3, and 0.88-0.90 [231-236/262] emulating PI-RADS ≥ 4 decisions.

STATISTICAL TESTS

Receiver operating characteristic (ROC) and area under the curve (AUC) was compared using DeLong and Obuchowski test. Sensitivity and specificity were compared using McNemar test. Statistical significance threshold was P = 0.05.

RESULTS

Test set (N = 695) ROC-AUC performance of SLA (trained with 200/500/998 exams) was 0.75/0.80/0.83, respectively. PLA achieved lower ROC-AUC of 0.64/0.72/0.78. Both increased performance significantly with increasing training set size. ROC-AUC for SLA at 500 exams was comparable to PLA at 998 exams (P = 0.28). ROC-AUC was significantly different between SLA and PLA at same training set sizes, however the ROC-AUC difference decreased significantly from 200 to 998 training exams. Emulating PI-RADS ≥ 3 decisions, difference between PLA specificity of 0.12 [51/433] and SLA specificity of 0.13 [55/433] became undetectable (P = 1.0) at 998 exams. Emulating PI-RADS ≥ 4 decisions, at 998 exams, SLA specificity of 0.51 [221/433] remained higher than PLA specificity at 0.39 [170/433]. However, PLA specificity at 998 exams became comparable to SLA specificity of 0.37 [159/433] at 200 exams (P = 0.70).

DATA CONCLUSION

Weakly supervised training of a classification CNN using patient-level-only annotation had lower performance compared to training with slice-wise annotations, but improved significantly faster with additional training data.

EVIDENCE LEVEL

3 TECHNICAL EFFICACY: Stage 2.

Deep-learning-based reconstruction of undersampled MRI to reduce scan times: a multicentre, retrospective, cohort study

BACKGROUND

The extended acquisition times required for MRI limit its availability in resource-constrained settings. Consequently, accelerating MRI by undersampling k-space data, which is necessary to reconstruct an image, has been a long-standing but important challenge. We aimed to develop a deep convolutional neural network (dCNN) optimisation method for MRI reconstruction and to reduce scan times and evaluate its effect on image quality and accuracy of oncological imaging biomarkers.

METHODS

In this multicentre, retrospective, cohort study, MRI data from patients with glioblastoma treated at Heidelberg University Hospital (775 patients and 775 examinations) and from the phase 2 CORE trial (260 patients, 1083 examinations, and 58 institutions) and the phase 3 CENTRIC trial (505 patients, 3147 examinations, and 139 institutions) were used to develop, train, and test dCNN for reconstructing MRI from highly undersampled single-coil k-space data with various acceleration rates (R=2, 4, 6, 8, 10, and 15). Independent testing was performed with MRIs from the phase 2/3 EORTC-26101 trial (528 patients with glioblastoma, 1974 examinations, and 32 institutions). The similarity between undersampled dCNN-reconstructed and original MRIs was quantified with various image quality metrics, including structural similarity index measure (SSIM) and the accuracy of undersampled dCNN-reconstructed MRI on downstream radiological assessment of imaging biomarkers in oncology (automated artificial intelligence-based quantification of tumour burden and treatment response) was performed in the EORTC-26101 test dataset. The public NYU Langone Health fastMRI brain test dataset (558 patients and 558 examinations) was used to validate the generalisability and robustness of the dCNN for reconstructing MRIs from available multi-coil (parallel imaging) k-space data.

FINDINGS

In the EORTC-26101 test dataset, the median SSIM of undersampled dCNN-reconstructed MRI ranged from 0·88 to 0·99 across different acceleration rates, with 0·92 (95% CI 0·92-0·93) for 10-times acceleration (R=10). The 10-times undersampled dCNN-reconstructed MRI yielded excellent agreement with original MRI when assessing volumes of contrast-enhancing tumour (median DICE for spatial agreement of 0·89 [95% CI 0·88 to 0·89]; median volume difference of 0·01 cm3 [95% CI 0·00 to 0·03] equalling 0·21%; p=0·0036 for equivalence) or non-enhancing tumour or oedema (median DICE of 0·94 [95% CI 0·94 to 0·95]; median volume difference of -0·79 cm3 [95% CI -0·87 to -0·72] equalling -1·77%; p=0·023 for equivalence) in the EORTC-26101 test dataset. Automated volumetric tumour response assessment in the EORTC-26101 test dataset yielded an identical median time to progression of 4·27 months (95% CI 4·14 to 4·57) when using 10-times-undersampled dCNN-reconstructed or original MRI (log-rank p=0·80) and agreement in the time to progression in 374 (95·2%) of 393 patients with data. The dCNN generalised well to the fastMRI brain dataset, with significant improvements in the median SSIM when using multi-coil compared with single-coil k-space data (p<0·0001).

INTERPRETATION

Deep-learning-based reconstruction of undersampled MRI allows for a substantial reduction of scan times, with a 10-times acceleration demonstrating excellent image quality while preserving the accuracy of derived imaging biomarkers for the assessment of oncological treatment response. Our developments are available as open source software and hold considerable promise for increasing the accessibility to MRI, pending further prospective validation.

FUNDING

Deutsche Forschungsgemeinschaft (German Research Foundation) and an Else Kröner Clinician Scientist Endowed Professorship by the Else Kröner Fresenius Foundation.

Microstructural white matter biomarkers of symptom severity and therapy outcome in catatonia: Rationale, study design and preliminary clinical data of the whiteCAT study

The number of magnetic resonance imaging (MRI) studies on neuronal correlates of catatonia has dramatically increased in the last 10 years, but conclusive findings on white matter (WM) tracts alterations underlying catatonic symptoms are still lacking. Therefore, we conduct an interdisciplinary longitudinal MRI study (whiteCAT) with two main objectives: First, we aim to enroll 100 psychiatric patients with and 50 psychiatric patients without catatonia according to ICD-11 who will undergo a deep phenotyping approach with an extensive battery of demographic, psychopathological, psychometric, neuropsychological, instrumental and diffusion MRI assessments at baseline and 12 weeks follow-up. So far, 28 catatonia patients and 40 patients with schizophrenia or other primary psychotic disorders or mood disorders without catatonia have been studied cross-sectionally. 49 out of 68 patients have completed longitudinal assessment, so far. Second, we seek to develop and implement a new method for semi-automatic fiber tract delineation using active learning. By training supportive machine learning algorithms on the fly that are custom tailored to the respective analysis pipeline used to obtain the tractogram as well as the WM tract of interest, we plan to streamline and speed up this tedious and error-prone task while at the same time increasing reproducibility and robustness of the extraction process. The goal is to develop robust neuroimaging biomarkers of symptom severity and therapy outcome based on WM tracts underlying catatonia. If our MRI study is successful, it will be the largest longitudinal study to date that has investigated WM tracts in catatonia patients.

Addressing image misalignments in multi-parametric prostate MRI for enhanced computer-aided diagnosis of prostate cancer

Prostate cancer (PCa) diagnosis on multi-parametric magnetic resonance images (MRI) requires radiologists with a high level of expertise. Misalignments between the MRI sequences can be caused by patient movement, elastic soft-tissue deformations, and imaging artifacts. They further increase the complexity of the task prompting radiologists to interpret the images. Recently, computer-aided diagnosis (CAD) tools have demonstrated potential for PCa diagnosis typically relying on complex co-registration of the input modalities. However, there is no consensus among research groups on whether CAD systems profit from using registration. Furthermore, alternative strategies to handle multi-modal misalignments have not been explored so far. Our study introduces and compares different strategies to cope with image misalignments and evaluates them regarding to their direct effect on diagnostic accuracy of PCa. In addition to established registration algorithms, we propose 'misalignment augmentation' as a concept to increase CAD robustness. As the results demonstrate, misalignment augmentations can not only compensate for a complete lack of registration, but if used in conjunction with registration, also improve the overall performance on an independent test set.

Application of a validated prostate MRI deep learning system to independent same-vendor multi-institutional data: demonstration of transferability

OBJECTIVES

To evaluate a fully automatic deep learning system to detect and segment clinically significant prostate cancer (csPCa) on same-vendor prostate MRI from two different institutions not contributing to training of the system.

MATERIALS AND METHODS

In this retrospective study, a previously bi-institutionally validated deep learning system (UNETM) was applied to bi-parametric prostate MRI data from one external institution (A), a PI-RADS distribution-matched internal cohort (B), and a csPCa stratified subset of single-institution external public challenge data (C). csPCa was defined as ISUP Grade Group ≥ 2 determined from combined targeted and extended systematic MRI/transrectal US-fusion biopsy. Performance of UNETM was evaluated by comparing ROC AUC and specificity at typical PI-RADS sensitivity levels. Lesion-level analysis between UNETM segmentations and radiologist-delineated segmentations was performed using Dice coefficient, free-response operating characteristic (FROC), and weighted alternative (waFROC). The influence of using different diffusion sequences was analyzed in cohort A.

RESULTS

In 250/250/140 exams in cohorts A/B/C, differences in ROC AUC were insignificant with 0.80 (95% CI: 0.74-0.85)/0.87 (95% CI: 0.83-0.92)/0.82 (95% CI: 0.75-0.89). At sensitivities of 95% and 90%, UNETM achieved specificity of 30%/50% in A, 44%/71% in B, and 43%/49% in C, respectively. Dice coefficient of UNETM and radiologist-delineated lesions was 0.36 in A and 0.49 in B. The waFROC AUC was 0.67 (95% CI: 0.60-0.83) in A and 0.7 (95% CI: 0.64-0.78) in B. UNETM performed marginally better on readout-segmented than on single-shot echo-planar-imaging.

CONCLUSION

For same-vendor examinations, deep learning provided comparable discrimination of csPCa and non-csPCa lesions and examinations between local and two independent external data sets, demonstrating the applicability of the system to institutions not participating in model training.

CLINICAL RELEVANCE STATEMENT

A previously bi-institutionally validated fully automatic deep learning system maintained acceptable exam-level diagnostic performance in two independent external data sets, indicating the potential of deploying AI models without retraining or fine-tuning, and corroborating evidence that AI models extract a substantial amount of transferable domain knowledge about MRI-based prostate cancer assessment.

KEY POINTS

• A previously bi-institutionally validated fully automatic deep learning system maintained acceptable exam-level diagnostic performance in two independent external data sets. • Lesion detection performance and segmentation congruence was similar on the institutional and an external data set, as measured by the weighted alternative FROC AUC and Dice coefficient. • Although the system generalized to two external institutions without re-training, achieving expected sensitivity and specificity levels using the deep learning system requires probability thresholds to be adjusted, underlining the importance of institution-specific calibration and quality control.

The Cell Tracking Challenge: 10 years of objective benchmarking

The Cell Tracking Challenge is an ongoing benchmarking initiative that has become a reference in cell segmentation and tracking algorithm development. Here, we present a significant number of improvements introduced in the challenge since our 2017 report. These include the creation of a new segmentation-only benchmark, the enrichment of the dataset repository with new datasets that increase its diversity and complexity, and the creation of a silver standard reference corpus based on the most competitive results, which will be of particular interest for data-hungry deep learning-based strategies. Furthermore, we present the up-to-date cell segmentation and tracking leaderboards, an in-depth analysis of the relationship between the performance of the state-of-the-art methods and the properties of the datasets and annotations, and two novel, insightful studies about the generalizability and the reusability of top-performing methods. These studies provide critical practical conclusions for both developers and users of traditional and machine learning-based cell segmentation and tracking algorithms.

Analysis of volume and topography of adipose tissue in the trunk: Results of MRI of 11,141 participants in the German National Cohort

This research addresses the assessment of adipose tissue (AT) and spatial distribution of visceral (VAT) and subcutaneous fat (SAT) in the trunk from standardized magnetic resonance imaging at 3 T, thereby demonstrating the feasibility of deep learning (DL)-based image segmentation in a large population-based cohort in Germany (five sites). Volume and distribution of AT play an essential role in the pathogenesis of insulin resistance, a risk factor of developing metabolic/cardiovascular diseases. Cross-validated training of the DL-segmentation model led to a mean Dice similarity coefficient of >0.94, corresponding to a mean absolute volume deviation of about 22 ml. SAT is significantly increased in women compared to men, whereas VAT is increased in males. Spatial distribution shows age- and body mass index-related displacements. DL-based image segmentation provides robust and fast quantification of AT (≈15 s per dataset versus 3 to 4 hours for manual processing) and assessment of its spatial distribution from magnetic resonance images in large cohort studies.

Automated lung cancer assessment on 18F-PET/CT using Retina U-Net and anatomical region segmentation

OBJECTIVES

To develop and test a Retina U-Net algorithm for the detection of primary lung tumors and associated metastases of all stages on FDG-PET/CT.

METHODS

A data set consisting of 364 FDG-PET/CTs of patients with histologically confirmed lung cancer was used for algorithm development and internal testing. The data set comprised tumors of all stages. All lung tumors (T), lymphatic metastases (N), and distant metastases (M) were manually segmented as 3D volumes using whole-body PET/CT series. The data set was split into a training (n = 216), validation (n = 74), and internal test data set (n = 74). Detection performance for all lesion types at multiple classifier thresholds was evaluated and false-positive-findings-per-case (FP/c) calculated. Next, detected lesions were assigned to categories T, N, or M using an automated anatomical region segmentation. Furthermore, reasons for FPs were visually assessed and analyzed. Finally, performance was tested on 20 PET/CTs from another institution.

RESULTS

Sensitivity for T lesions was 86.2% (95% CI: 77.2-92.7) at a FP/c of 2.0 on the internal test set. The anatomical correlate to most FPs was the physiological activity of bone marrow (16.8%). TNM categorization based on the anatomical region approach was correct in 94.3% of lesions. Performance on the external test set confirmed the good performance of the algorithm (overall detection rate = 88.8% (95% CI: 82.5-93.5%) and FP/c = 2.7).

CONCLUSIONS

Retina U-Nets are a valuable tool for tumor detection tasks on PET/CT and can form the backbone of reading assistance tools in this field. FPs have anatomical correlates that can lead the way to further algorithm improvements. The code is publicly available.

KEY POINTS

• Detection of malignant lesions in PET/CT with Retina U-Net is feasible. • All false-positive findings had anatomical correlates, physiological bone marrow activity being the most prevalent. • Retina U-Nets can build the backbone for tools assisting imaging professionals in lung tumor staging.

Rapid artificial intelligence solutions in a pandemic-The COVID-19-20 Lung CT Lesion Segmentation Challenge

Artificial intelligence (AI) methods for the automatic detection and quantification of COVID-19 lesions in chest computed tomography (CT) might play an important role in the monitoring and management of the disease. We organized an international challenge and competition for the development and comparison of AI algorithms for this task, which we supported with public data and state-of-the-art benchmark methods. Board Certified Radiologists annotated 295 public images from two sources (A and B) for algorithms training (n=199, source A), validation (n=50, source A) and testing (n=23, source A; n=23, source B). There were 1,096 registered teams of which 225 and 98 completed the validation and testing phases, respectively. The challenge showed that AI models could be rapidly designed by diverse teams with the potential to measure disease or facilitate timely and patient-specific interventions. This paper provides an overview and the major outcomes of the COVID-19 Lung CT Lesion Segmentation Challenge — 2020.

Band selection for oxygenation estimation with multispectral/hyperspectral imaging

Multispectral imaging provides valuable information on tissue composition such as hemoglobin oxygen saturation. However, the real-time application of this technique in interventional medicine can be challenging due to the long acquisition times needed for large amounts of hyperspectral data with hundreds of bands. While this challenge can partially be addressed by choosing a discriminative subset of bands, the band selection methods proposed to date are mainly restricted by the availability of often hard to obtain reference measurements. We address this bottleneck with a new approach to band selection that leverages highly accurate Monte Carlo (MC) simulations. We hypothesize that a so chosen small subset of bands can reproduce or even improve upon the results of a quasi continuous spectral measurement. We further investigate whether novel domain adaptation techniques can address the inevitable domain shift stemming from the use of simulations. Initial results based on in silico and in vivo experiments suggest that 10-20 bands are sufficient to closely reproduce results from spectral measurements with 101 bands in the 500-700 nm range. The investigated domain adaptation technique, which only requires unlabeled in vivo measurements, yielded better results than the pure in silico band selection method. Overall, our method could guide development of fast multispectral imaging systems suited for interventional use without relying on complex hardware setups or manually labeled data.

Automated detection and quantification of brain metastases on clinical MRI data using artificial neural networks

Background

Reliable detection and precise volumetric quantification of brain metastases (BM) on MRI are essential for guiding treatment decisions. Here we evaluate the potential of artificial neural networks (ANN) for automated detection and quantification of BM.

Methods

A consecutive series of 308 patients with BM was used for developing an ANN (with a 4:1 split for training/testing) for automated volumetric assessment of contrast-enhancing tumors (CE) and non-enhancing FLAIR signal abnormality including edema (NEE). An independent consecutive series of 30 patients was used for external testing. Performance was assessed case-wise for CE and NEE and lesion-wise for CE using the case-wise/lesion-wise DICE-coefficient (C/L-DICE), positive predictive value (L-PPV) and sensitivity (C/L-Sensitivity).

Results

The performance of detecting CE lesions on the validation dataset was not significantly affected when evaluating different volumetric thresholds (0.001–0.2 cm3; P = .2028). The median L-DICE and median C-DICE for CE lesions were 0.78 (IQR = 0.6–0.91) and 0.90 (IQR = 0.85–0.94) in the institutional as well as 0.79 (IQR = 0.67–0.82) and 0.84 (IQR = 0.76–0.89) in the external test dataset. The corresponding median L-Sensitivity and median L-PPV were 0.81 (IQR = 0.63–0.92) and 0.79 (IQR = 0.63–0.93) in the institutional test dataset, as compared to 0.85 (IQR = 0.76–0.94) and 0.76 (IQR = 0.68–0.88) in the external test dataset. The median C-DICE for NEE was 0.96 (IQR = 0.92–0.97) in the institutional test dataset as compared to 0.85 (IQR = 0.72–0.91) in the external test dataset.

Conclusion

The developed ANN-based algorithm (publicly available at www.github.com/NeuroAI-HD/HD-BM) allows reliable detection and precise volumetric quantification of CE and NEE compartments in patients with BM.

Fully Automatic Deep Learning in Bi-institutional Prostate Magnetic Resonance Imaging: Effects of Cohort Size and Heterogeneity

BACKGROUND

The potential of deep learning to support radiologist prostate magnetic resonance imaging (MRI) interpretation has been demonstrated.

PURPOSE

The aim of this study was to evaluate the effects of increased and diversified training data (TD) on deep learning performance for detection and segmentation of clinically significant prostate cancer-suspicious lesions.

MATERIALS AND METHODS

In this retrospective study, biparametric (T2-weighted and diffusion-weighted) prostate MRI acquired with multiple 1.5-T and 3.0-T MRI scanners in consecutive men was used for training and testing of prostate segmentation and lesion detection networks. Ground truth was the combination of targeted and extended systematic MRI-transrectal ultrasound fusion biopsies, with significant prostate cancer defined as International Society of Urological Pathology grade group greater than or equal to 2. U-Nets were internally validated on full, reduced, and PROSTATEx-enhanced training sets and subsequently externally validated on the institutional test set and the PROSTATEx test set. U-Net segmentation was calibrated to clinically desired levels in cross-validation, and test performance was subsequently compared using sensitivities, specificities, predictive values, and Dice coefficient.

RESULTS

One thousand four hundred eighty-eight institutional examinations (median age, 64 years; interquartile range, 58-70 years) were temporally split into training (2014-2017, 806 examinations, supplemented by 204 PROSTATEx examinations) and test (2018-2020, 682 examinations) sets. In the test set, Prostate Imaging-Reporting and Data System (PI-RADS) cutoffs greater than or equal to 3 and greater than or equal to 4 on a per-patient basis had sensitivity of 97% (241/249) and 90% (223/249) at specificity of 19% (82/433) and 56% (242/433), respectively. The full U-Net had corresponding sensitivity of 97% (241/249) and 88% (219/249) with specificity of 20% (86/433) and 59% (254/433), not statistically different from PI-RADS (P > 0.3 for all comparisons). U-Net trained using a reduced set of 171 consecutive examinations achieved inferior performance (P < 0.001). PROSTATEx training enhancement did not improve performance. Dice coefficients were 0.90 for prostate and 0.42/0.53 for MRI lesion segmentation at PI-RADS category 3/4 equivalents.

CONCLUSIONS

In a large institutional test set, U-Net confirms similar performance to clinical PI-RADS assessment and benefits from more TD, with neither institutional nor PROSTATEx performance improved by adding multiscanner or bi-institutional TD.

Deep-learning-based synthesis of post-contrast T1-weighted MRI for tumour response assessment in neuro-oncology: a multicentre, retrospective cohort study

BACKGROUND

Gadolinium-based contrast agents (GBCAs) are widely used to enhance tissue contrast during MRI scans and play a crucial role in the management of patients with cancer. However, studies have shown gadolinium deposition in the brain after repeated GBCA administration with yet unknown clinical significance. We aimed to assess the feasibility and diagnostic value of synthetic post-contrast T1-weighted MRI generated from pre-contrast MRI sequences through deep convolutional neural networks (dCNN) for tumour response assessment in neuro-oncology.

METHODS

In this multicentre, retrospective cohort study, we used MRI examinations to train and validate a dCNN for synthesising post-contrast T1-weighted sequences from pre-contrast T1-weighted, T2-weighted, and fluid-attenuated inversion recovery sequences. We used MRI scans with availability of these sequences from 775 patients with glioblastoma treated at Heidelberg University Hospital, Heidelberg, Germany (775 MRI examinations); 260 patients who participated in the phase 2 CORE trial (1083 MRI examinations, 59 institutions); and 505 patients who participated in the phase 3 CENTRIC trial (3147 MRI examinations, 149 institutions). Separate training runs to rank the importance of individual sequences and (for a subset) diffusion-weighted imaging were conducted. Independent testing was performed on MRI data from the phase 2 and phase 3 EORTC-26101 trial (521 patients, 1924 MRI examinations, 32 institutions). The similarity between synthetic and true contrast enhancement on post-contrast T1-weighted MRI was quantified using the structural similarity index measure (SSIM). Automated tumour segmentation and volumetric tumour response assessment based on synthetic versus true post-contrast T1-weighted sequences was performed in the EORTC-26101 trial and agreement was assessed with Kaplan-Meier plots.

FINDINGS

The median SSIM score for predicting contrast enhancement on synthetic post-contrast T1-weighted sequences in the EORTC-26101 test set was 0·818 (95% CI 0·817-0·820). Segmentation of the contrast-enhancing tumour from synthetic post-contrast T1-weighted sequences yielded a median tumour volume of 6·31 cm³ (5·60 to 7·14), thereby underestimating the true tumour volume by a median of -0·48 cm³ (-0·37 to -0·76) with the concordance correlation coefficient suggesting a strong linear association between tumour volumes derived from synthetic versus true post-contrast T1-weighted sequences (0·782, 0·751-0·807, p<0·0001). Volumetric tumour response assessment in the EORTC-26101 trial showed a median time to progression of 4·2 months (95% CI 4·1-5·2) with synthetic post-contrast T1-weighted and 4·3 months (4·1-5·5) with true post-contrast T1-weighted sequences (p=0·33). The strength of the association between the time to progression as a surrogate endpoint for predicting the patients' overall survival in the EORTC-26101 cohort was similar when derived from synthetic post-contrast T1-weighted sequences (hazard ratio of 1·749, 95% CI 1·282-2·387, p=0·0004) and model C-index (0·667, 0·622-0·708) versus true post-contrast T1-weighted MRI (1·799, 95% CI 1·314-2·464, p=0·0003) and model C-index (0·673, 95% CI 0·626-0·711).

INTERPRETATION

Generating synthetic post-contrast T1-weighted MRI from pre-contrast MRI using dCNN is feasible and quantification of the contrast-enhancing tumour burden from synthetic post-contrast T1-weighted MRI allows assessment of the patient's response to treatment with no significant difference by comparison with true post-contrast T1-weighted sequences with administration of GBCAs. This finding could guide the application of dCNN in radiology to potentially reduce the necessity of GBCA administration.

FUNDING

Deutsche Forschungsgemeinschaft.

White matter microstructure alterations in cortico-striatal networks are associated with parkinsonism in schizophrenia spectrum disorders

The specific role of white matter (WM) microstructure in parkinsonism among patients with schizophrenia spectrum disorders (SSD) is largely unknown. To determine whether topographical alterations of WM microstructure contribute to parkinsonism in SSD patients, we examined healthy controls (HC, n=16) and SSD patients with and without parkinsonism, as defined by Simpson-Angus Scale total score of ≥4 (SSD-P, n=33) or <4 (SSD-nonP, n=62). We used whole brain tract-based spatial statistics (TBSS), tractometry (along tract statistics using TractSeg) and graph analytics (clustering coefficient (CCO), local betweenness centrality (BC)) to provide a framework of specific WM microstructural changes underlying parkinsonism in SSD. Using these methods, post hoc analyses showed (a) decreased fractional anisotrophy (FA), as measured via tractometry, in the corpus callosum, corticospinal tract and striato-fronto-orbital tract, and (b) increased CCO, as derived by graph analytics, in the left orbitofrontal cortex (OFC) and left superior frontal gyrus (SFG), in SSD-P patients when compared to SSD-nonP patients. Increased CCO in the left OFC and SFG was associated with SAS scores. These findings indicate the prominence of OFC alterations and aberrant connectivity with fronto-parietal regions and striatum in the pathogenesis of parkinsonism in SSD. This study further supports the notion of altered "bottom-up modulation" between basal ganglia and fronto-parietal regions in the pathobiology of parkinsonism, which may reflect an interaction between movement disorder intrinsic to SSD and antipsychotic drug-induced sensorimotor dysfunction.

Rapid Artificial Intelligence Solutions in a Pandemic - The COVID-19-20 Lung CT Lesion Segmentation Challenge

Artificial intelligence (AI) methods for the automatic detection and quantification of COVID-19 lesions in chest computed tomography (CT) might play an important role in the monitoring and management of the disease. We organized an international challenge and competition for the development and comparison of AI algorithms for this task, which we supported with public data and state-of-the-art benchmark methods. Board Certified Radiologists annotated 295 public images from two sources (A and B) for algorithms training (n=199, source A), validation (n=50, source A) and testing (n=23, source A; n=23, source B). There were 1,096 registered teams of which 225 and 98 completed the validation and testing phases, respectively. The challenge showed that AI models could be rapidly designed by diverse teams with the potential to measure disease or facilitate timely and patient-specific interventions. This paper provides an overview and the major outcomes of the COVID-19 Lung CT Lesion Segmentation Challenge - 2020.

Fully Automated Segmentation of Pulmonary Fibrosis Using Different Software Tools

OBJECTIVE

Evaluation of software tools for segmentation, quantification, and characterization of fibrotic pulmonary parenchyma changes will strengthen the role of CT as biomarkers of disease extent, evolution, and response to therapy in idiopathic pulmonary fibrosis (IPF) patients.

METHODS

418 nonenhanced thin-section MDCTs of 127 IPF patients and 78 MDCTs of 78 healthy individuals were analyzed through 3 fully automated, completely different software tools: YACTA, LUFIT, and IMBIO. The agreement between YACTA and LUFIT on segmented lung volume and 80th (reflecting fibrosis) and 40th (reflecting ground-glass opacity) percentile of the lung density histogram was analyzed using Bland-Altman plots. The fibrosis and ground-glass opacity segmented by IMBIO (lung texture analysis software tool) were included in specific regression analyses.

RESULTS

In the IPF-group, LUFIT outperformed YACTA by segmenting more lung volume (mean difference 242 mL, 95% limits of agreement -54 to 539 mL), as well as quantifying higher 80th (76 HU, -6 to 158 HU) and 40th percentiles (9 HU, -73 to 90 HU). No relevant differences were revealed in the control group. The 80th/40th percentile as quantified by LUFIT correlated positively with the percentage of fibrosis/ground-glass opacity calculated by IMBIO (r = 0.78/r = 0.92).

CONCLUSIONS

In terms of segmentation of pulmonary fibrosis, LUFIT as a shape model-based segmentation software tool is superior to the threshold-based YACTA, tool, since the density of (severe) fibrosis is similar to that of the surrounding soft tissues. Therefore, shape modeling as used in LUFIT may serve as a valid tool in the quantification of IPF, since this mainly affects the subpleural space.

Adapting Bidirectional Encoder Representations from Transformers (BERT) to Assess Clinical Semantic Textual Similarity: Algorithm Development and Validation Study

Background

Natural Language Understanding enables automatic extraction of relevant information from clinical text data, which are acquired every day in hospitals. In 2018, the language model Bidirectional Encoder Representations from Transformers (BERT) was introduced, generating new state-of-the-art results on several downstream tasks. The National NLP Clinical Challenges (n2c2) is an initiative that strives to tackle such downstream tasks on domain-specific clinical data. In this paper, we present the results of our participation in the 2019 n2c2 and related work completed thereafter.

Objective

The objective of this study was to optimally leverage BERT for the task of assessing the semantic textual similarity of clinical text data.

Methods

We used BERT as an initial baseline and analyzed the results, which we used as a starting point to develop 3 different approaches where we (1) added additional, handcrafted sentence similarity features to the classifier token of BERT and combined the results with more features in multiple regression estimators, (2) incorporated a built-in ensembling method, M-Heads, into BERT by duplicating the regression head and applying an adapted training strategy to facilitate the focus of the heads on different input patterns of the medical sentences, and (3) developed a graph-based similarity approach for medications, which allows extrapolating similarities across known entities from the training set. The approaches were evaluated with the Pearson correlation coefficient between the predicted scores and ground truth of the official training and test dataset.

Results

We improved the performance of BERT on the test dataset from a Pearson correlation coefficient of 0.859 to 0.883 using a combination of the M-Heads method and the graph-based similarity approach. We also show differences between the test and training dataset and how the two datasets influenced the results.

Conclusions

We found that using a graph-based similarity approach has the potential to extrapolate domain specific knowledge to unseen sentences. We observed that it is easily possible to obtain deceptive results from the test dataset, especially when the distribution of the data samples is different between training and test datasets.

The state of the art in kidney and kidney tumor segmentation in contrast-enhanced CT imaging: Results of the KiTS19 challenge

There is a large body of literature linking anatomic and geometric characteristics of kidney tumors to perioperative and oncologic outcomes. Semantic segmentation of these tumors and their host kidneys is a promising tool for quantitatively characterizing these lesions, but its adoption is limited due to the manual effort required to produce high-quality 3D segmentations of these structures. Recently, methods based on deep learning have shown excellent results in automatic 3D segmentation, but they require large datasets for training, and there remains little consensus on which methods perform best. The 2019 Kidney and Kidney Tumor Segmentation challenge (KiTS19) was a competition held in conjunction with the 2019 International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) which sought to address these issues and stimulate progress on this automatic segmentation problem. A training set of 210 cross sectional CT images with kidney tumors was publicly released with corresponding semantic segmentation masks. 106 teams from five continents used this data to develop automated systems to predict the true segmentation masks on a test set of 90 CT images for which the corresponding ground truth segmentations were kept private. These predictions were scored and ranked according to their average Sørensen-Dice coefficient between the kidney and tumor across all 90 cases. The winning team achieved a Dice of 0.974 for kidney and 0.851 for tumor, approaching the inter-annotator performance on kidney (0.983) but falling short on tumor (0.923). This challenge has now entered an "open leaderboard" phase where it serves as a challenging benchmark in 3D semantic segmentation.

CHAOS Challenge - combined (CT-MR) healthy abdominal organ segmentation

Segmentation of abdominal organs has been a comprehensive, yet unresolved, research field for many years. In the last decade, intensive developments in deep learning (DL) introduced new state-of-the-art segmentation systems. Despite outperforming the overall accuracy of existing systems, the effects of DL model properties and parameters on the performance are hard to interpret. This makes comparative analysis a necessary tool towards interpretable studies and systems. Moreover, the performance of DL for emerging learning approaches such as cross-modality and multi-modal semantic segmentation tasks has been rarely discussed. In order to expand the knowledge on these topics, the CHAOS - Combined (CT-MR) Healthy Abdominal Organ Segmentation challenge was organized in conjunction with the IEEE International Symposium on Biomedical Imaging (ISBI), 2019, in Venice, Italy. Abdominal organ segmentation from routine acquisitions plays an important role in several clinical applications, such as pre-surgical planning or morphological and volumetric follow-ups for various diseases. These applications require a certain level of performance on a diverse set of metrics such as maximum symmetric surface distance (MSSD) to determine surgical error-margin or overlap errors for tracking size and shape differences. Previous abdomen related challenges are mainly focused on tumor/lesion detection and/or classification with a single modality. Conversely, CHAOS provides both abdominal CT and MR data from healthy subjects for single and multiple abdominal organ segmentation. Five different but complementary tasks were designed to analyze the capabilities of participating approaches from multiple perspectives. The results were investigated thoroughly, compared with manual annotations and interactive methods. The analysis shows that the performance of DL models for single modality (CT / MR) can show reliable volumetric analysis performance (DICE: 0.98 ± 0.00 / 0.95 ± 0.01), but the best MSSD performance remains limited (21.89 ± 13.94 / 20.85 ± 10.63 mm). The performances of participating models decrease dramatically for cross-modality tasks both for the liver (DICE: 0.88 ± 0.15 MSSD: 36.33 ± 21.97 mm). Despite contrary examples on different applications, multi-tasking DL models designed to segment all organs are observed to perform worse compared to organ-specific ones (performance drop around 5%). Nevertheless, some of the successful models show better performance with their multi-organ versions. We conclude that the exploration of those pros and cons in both single vs multi-organ and cross-modality segmentations is poised to have an impact on further research for developing effective algorithms that would support real-world clinical applications. Finally, having more than 1500 participants and receiving more than 550 submissions, another important contribution of this study is the analysis on shortcomings of challenge organizations such as the effects of multiple submissions and peeking phenomenon.

nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation

Biomedical imaging is a driver of scientific discovery and a core component of medical care and is being stimulated by the field of deep learning. While semantic segmentation algorithms enable image analysis and quantification in many applications, the design of respective specialized solutions is non-trivial and highly dependent on dataset properties and hardware conditions. We developed nnU-Net, a deep learning-based segmentation method that automatically configures itself, including preprocessing, network architecture, training and post-processing for any new task. The key design choices in this process are modeled as a set of fixed parameters, interdependent rules and empirical decisions. Without manual intervention, nnU-Net surpasses most existing approaches, including highly specialized solutions on 23 public datasets used in international biomedical segmentation competitions. We make nnU-Net publicly available as an out-of-the-box tool, rendering state-of-the-art segmentation accessible to a broad audience by requiring neither expert knowledge nor computing resources beyond standard network training.

Comparative validation of multi-instance instrument segmentation in endoscopy: Results of the ROBUST-MIS 2019 challenge

Intraoperative tracking of laparoscopic instruments is often a prerequisite for computer and robotic-assisted interventions. While numerous methods for detecting, segmenting and tracking of medical instruments based on endoscopic video images have been proposed in the literature, key limitations remain to be addressed: Firstly, robustness, that is, the reliable performance of state-of-the-art methods when run on challenging images (e.g. in the presence of blood, smoke or motion artifacts). Secondly, generalization; algorithms trained for a specific intervention in a specific hospital should generalize to other interventions or institutions. In an effort to promote solutions for these limitations, we organized the Robust Medical Instrument Segmentation (ROBUST-MIS) challenge as an international benchmarking competition with a specific focus on the robustness and generalization capabilities of algorithms. For the first time in the field of endoscopic image processing, our challenge included a task on binary segmentation and also addressed multi-instance detection and segmentation. The challenge was based on a surgical data set comprising 10,040 annotated images acquired from a total of 30 surgical procedures from three different types of surgery. The validation of the competing methods for the three tasks (binary segmentation, multi-instance detection and multi-instance segmentation) was performed in three different stages with an increasing domain gap between the training and the test data. The results confirm the initial hypothesis, namely that algorithm performance degrades with an increasing domain gap. While the average detection and segmentation quality of the best-performing algorithms is high, future research should concentrate on detection and segmentation of small, crossing, moving and transparent instrument(s) (parts).

Validation of diffusion MRI phenotypes for predicting response to bevacizumab in recurrent glioblastoma: post-hoc analysis of the EORTC-26101 trial

BACKGROUND

This study validated a previously described diffusion MRI phenotype as a potential predictive imaging biomarker in patients with recurrent glioblastoma receiving bevacizumab (BEV).

METHODS

A total of 396/596 patients (66%) from the prospective randomized phase II/III EORTC-26101 trial (with n = 242 in the BEV and n = 154 in the non-BEV arm) met the inclusion criteria with availability of anatomical and diffusion MRI sequences at baseline prior treatment. Apparent diffusion coefficient (ADC) histograms from the contrast-enhancing tumor volume were fitted to a double Gaussian distribution and the mean of the lower curve (ADClow) was used for further analysis. The predictive ability of ADClow was assessed with biomarker threshold models and multivariable Cox regression for overall survival (OS) and progression-free survival (PFS).

RESULTS

ADClow was associated with PFS (hazard ratio [HR] = 0.625, P = 0.007) and OS (HR = 0.656, P = 0.031). However, no (predictive) interaction between ADClow and the treatment arm was present (P = 0.865 for PFS, P = 0.722 for OS). Independent (prognostic) significance of ADClow was retained after adjusting for epidemiological, clinical, and molecular characteristics (P ≤ 0.02 for OS, P ≤ 0.01 PFS). The biomarker threshold model revealed an optimal ADClow cutoff of 1241*10-6 mm2/s for OS. Thereby, median OS for BEV-patients with ADClow ≥ 1241 was 10.39 months versus 8.09 months for those with ADClow < 1241 (P = 0.004). Similarly, median OS for non-BEV patients with ADClow ≥ 1241 was 9.80 months versus 7.79 months for those with ADClow < 1241 (P = 0.054).

CONCLUSIONS

ADClow is an independent prognostic parameter for stratifying OS and PFS in patients with recurrent glioblastoma. Consequently, the previously suggested role of ADClow as predictive imaging biomarker could not be confirmed within this phase II/III trial.

Multiparametric mapping of white matter microstructure in catatonia

Catatonia is characterized by motor, affective and behavioral abnormalities. To date, the specific role of white matter (WM) abnormalities in schizophrenia spectrum disorders (SSD) patients with catatonia is largely unknown. In this study, diffusion magnetic resonance imaging (dMRI) data were collected from 111 right-handed SSD patients and 28 healthy controls. Catatonic symptoms were examined on the Northoff Catatonia Rating Scale (NCRS). We used whole-brain tract-based spatial statistics (TBSS), tractometry (along tract statistics using TractSeg) and graph analytics (clustering coefficient-CCO, local betweenness centrality-BC) to provide a framework of specific WM microstructural abnormalities underlying catatonia in SSD. Following a categorical approach, post hoc analyses showed differences in fractional anisotrophy (FA) measured via tractometry in the corpus callosum, corticospinal tract and thalamo-premotor tract as well as increased CCO as derived by graph analytics of the right superior parietal cortex (SPC) and left caudate nucleus in catatonic patients (NCRS total score ≥ 3; n = 30) when compared to non-catatonic patients (NCRS total score = 0; n = 29). In catatonic patients according to DSM-IV-TR (n = 43), catatonic symptoms were associated with FA variations (tractometry) of the left corticospinal tract and CCO of the left orbitofrontal cortex, primary motor cortex, supplementary motor area and putamen. This study supports the notion that structural reorganization of WM bundles connecting orbitofrontal/parietal, thalamic and striatal regions contribute to catatonia in SSD patients.

Simulated clinical deployment of fully automatic deep learning for clinical prostate MRI assessment

Objectives

To simulate clinical deployment, evaluate performance, and establish quality assurance of a deep learning algorithm (U-Net) for detection, localization, and segmentation of clinically significant prostate cancer (sPC), ISUP grade group ≥ 2, using bi-parametric MRI.

Methods

In 2017, 284 consecutive men in active surveillance, biopsy-naïve or pre-biopsied, received targeted and extended systematic MRI/transrectal US-fusion biopsy, after examination on a single MRI scanner (3 T). A prospective adjustment scheme was evaluated comparing the performance of the Prostate Imaging Reporting and Data System (PI-RADS) and U-Net using sensitivity, specificity, predictive values, and the Dice coefficient.

Results

In the 259 eligible men (median 64 [IQR 61–72] years), PI-RADS had a sensitivity of 98% [106/108]/84% [91/108] with a specificity of 17% [25/151]/58% [88/151], for thresholds at ≥ 3/≥ 4 respectively. U-Net using dynamic threshold adjustment had a sensitivity of 99% [107/108]/83% [90/108] (p > 0.99/> 0.99) with a specificity of 24% [36/151]/55% [83/151] (p > 0.99/> 0.99) for probability thresholds d3 and d4 emulating PI-RADS ≥ 3 and ≥ 4 decisions respectively, not statistically different from PI-RADS. Co-occurrence of a radiological PI-RADS ≥ 4 examination and U-Net ≥ d3 assessment significantly improved the positive predictive value from 59 to 63% (p = 0.03), on a per-patient basis.

Conclusions

U-Net has similar performance to PI-RADS in simulated continued clinical use. Regular quality assurance should be implemented to ensure desired performance.

Key Points

• U-Net maintained similar diagnostic performance compared to radiological assessment of PI-RADS ≥ 4 when applied in a simulated clinical deployment.

• Application of our proposed prospective dynamic calibration method successfully adjusted U-Net performance within acceptable limits of the PI-RADS reference over time, while not being limited to PI-RADS as a reference.

• Simultaneous detection by U-Net and radiological assessment significantly improved the positive predictive value on a per-patient and per-lesion basis, while the negative predictive value remained unchanged.

Electronic supplementary material

The online version of this article (10.1007/s00330-020-07086-z) contains supplementary material, which is available to authorized users.

Noninvasive Characterization of Tumor Angiogenesis and Oxygenation in Bevacizumab-treated Recurrent Glioblastoma by Using Dynamic Susceptibility MRI: Secondary Analysis of the European Organization for Research and Treatment of Cancer 26101 Trial

Background Relevance of antiangiogenic treatment with bevacizumab in patients with glioblastoma is controversial because progression-free survival benefit did not translate into an overall survival (OS) benefit in randomized phase III trials. Purpose To perform longitudinal characterization of intratumoral angiogenesis and oxygenation by using dynamic susceptibility contrast agent-enhanced (DSC) MRI and evaluate its potential for predicting outcome from administration of bevacizumab. Materials and Methods In this secondary analysis of the prospective randomized phase II/III European Organization for Research and Treatment of Cancer 26101 trial conducted between October 2011 and December 2015 in 596 patients with first recurrence of glioblastoma, the subset of patients with availability of anatomic MRI and DSC MRI at baseline and first follow-up was analyzed. Patients were allocated into those administered bevacizumab (hereafter, the BEV group; either bevacizumab monotherapy or bevacizumab with lomustine) and those not administered bevacizumab (hereafter, the non-BEV group with lomustine monotherapy). Contrast-enhanced tumor volume, noncontrast-enhanced T2 fluid-attenuated inversion recovery (FLAIR) signal abnormality volume, Gaussian-normalized relative cerebral blood volume (nrCBV), Gaussian-normalized relative blood flow (nrCBF), and tumor metabolic rate of oxygen (nTMRO₂) was quantified. The predictive ability of these imaging parameters was assessed with multivariable Cox regression and formal interaction testing. Results A total of 254 of 596 patients were evaluated (mean age, 57 years ± 11; 155 men; 161 in the BEV group and 93 in non-BEV group). Progression-free survival was longer in the BEV group (3.7 months; 95% confidence interval [CI]: 3.0, 4.2) compared with the non-BEV group (2.5 months; 95% CI: 1.5, 2.9; P = .01), whereas OS was not different (P = .15). The nrCBV decreased for the BEV group (-16.3%; interquartile range [IQR], -39.5% to 12.0%; P = .01), but not for the non-BEV group (1.2%; IQR, -17.9% to 23.3%; P = .19) between baseline and first follow-up. An identical pattern was observed for both nrCBF and nTMRO₂ values. Contrast-enhanced tumor and noncontrast-enhanced T2 FLAIR signal abnormality volumes decreased for the BEV group (-66% [IQR, -83% to -35%] and -33% [IQR, -71% to -5%], respectively; P < .001 for both), whereas they increased for the non-BEV group (30% [IQR, -17% to 98%], P = .001; and 10% [IQR, -13% to 82%], P = .02, respectively) between baseline and first follow-up. None of the assessed MRI parameters were predictive for OS in the BEV group. Conclusion Bevacizumab treatment decreased tumor volumes, angiogenesis, and oxygenation, thereby reflecting its effectiveness for extending progression-free survival; however, these parameters were not predictive of overall survival (OS), which highlighted the challenges of identifying patients that derive an OS benefit from bevacizumab. © RSNA, 2020 Online supplemental material is available for this article. See also the editorial by Dillon in this issue.

The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping

Background Radiomic features may quantify characteristics present in medical imaging. However, the lack of standardized definitions and validated reference values have hampered clinical use. Purpose To standardize a set of 174 radiomic features. Materials and Methods Radiomic features were assessed in three phases. In phase I, 487 features were derived from the basic set of 174 features. Twenty-five research teams with unique radiomics software implementations computed feature values directly from a digital phantom, without any additional image processing. In phase II, 15 teams computed values for 1347 derived features using a CT image of a patient with lung cancer and predefined image processing configurations. In both phases, consensus among the teams on the validity of tentative reference values was measured through the frequency of the modal value and classified as follows: less than three matches, weak; three to five matches, moderate; six to nine matches, strong; 10 or more matches, very strong. In the final phase (phase III), a public data set of multimodality images (CT, fluorine 18 fluorodeoxyglucose PET, and T1-weighted MRI) from 51 patients with soft-tissue sarcoma was used to prospectively assess reproducibility of standardized features. Results Consensus on reference values was initially weak for 232 of 302 features (76.8%) at phase I and 703 of 1075 features (65.4%) at phase II. At the final iteration, weak consensus remained for only two of 487 features (0.4%) at phase I and 19 of 1347 features (1.4%) at phase II. Strong or better consensus was achieved for 463 of 487 features (95.1%) at phase I and 1220 of 1347 features (90.6%) at phase II. Overall, 169 of 174 features were standardized in the first two phases. In the final validation phase (phase III), most of the 169 standardized features could be excellently reproduced (166 with CT; 164 with PET; and 164 with MRI). Conclusion A set of 169 radiomics features was standardized, which enabled verification and calibration of different radiomics software. © RSNA, 2020 Online supplemental material is available for this article. See also the editorial by Kuhl and Truhn in this issue.

Exploring cortical predictors of clinical response to electroconvulsive therapy in major depression

Electroconvulsive therapy (ECT) is a rapid and highly effective treatment option for treatment-resistant major depressive disorder (TRD). The neural mechanisms underlying such beneficial effects are poorly understood. Exploring associations between changes of brain structure and clinical response is crucial for understanding ECT mechanisms of action and relevant for the validation of potential biomarkers that can facilitate the prediction of ECT efficacy. The aim of this explorative study was to identify cortical predictors of clinical response in TRD patients treated with ECT. We longitudinally investigated 12 TRD patients before and after ECT. Twelve matched healthy controls were studied cross sectionally. Demographical, clinical, and structural magnetic resonance imaging data at 3 T and multiple cortical markers derived from surface-based morphometry (SBM) analyses were considered. Multiple regression models were computed to identify predictors of clinical response to ECT, as reflected by Hamilton Depression Rating Scale (HAMD) score changes. Symptom severity differences pre-post-ECT were predicted by models including demographic data, clinical data and SBM of frontal, cingulate, and entorhinal structures. Using all-subsets regression, a model comprising HAMD score at baseline and cortical thickness of the left rostral anterior cingulate gyrus explained most variance in the data (multiple R² = 0.82). The data suggest that SBM provides powerful measures for identifying biomarkers for ECT response in TRD. Rostral anterior cingulate thickness and HAMD score at baseline showed the greatest predictive power of clinical response, in contrast to cortical complexity, cortical gyrification, or demographical data.

Optimal Statistical Incorporation of Independent Feature Stability Information into Radiomics Studies

Conducting side experiments termed robustness experiments, to identify features that are stable with respect to rescans, annotation, or other confounding effects is an important element in radiomics research. However, the matter of how to include the finding of these experiments into the model building process still needs to be explored. Three different methods for incorporating prior knowledge into a radiomics modelling process were evaluated: the naïve approach (ignoring feature quality), the most common approach consisting of removing unstable features, and a novel approach using data augmentation for information transfer (DAFIT). Multiple experiments were conducted using both synthetic and publicly available real lung imaging patient data. Ignoring additional information from side experiments resulted in significantly overestimated model performances meaning the estimated mean area under the curve achieved with a model was increased. Removing unstable features improved the performance estimation, while slightly decreasing the model performance, i.e. decreasing the area under curve achieved with the model. The proposed approach was superior both in terms of the estimation of the model performance and the actual model performance. Our experiments show that data augmentation can prevent biases in performance estimation and has several advantages over the plain omission of the unstable feature. The actual gain that can be obtained depends on the quality and applicability of the prior information on the features in the given domain. This will be an important topic of future research.

Classification of Cancer at Prostate MRI: Deep Learning versus Clinical PI-RADS Assessment

Background Men suspected of having clinically significant prostate cancer (sPC) increasingly undergo prostate MRI. The potential of deep learning to provide diagnostic support for human interpretation requires further evaluation. Purpose To compare the performance of clinical assessment to a deep learning system optimized for segmentation trained with T2-weighted and diffusion MRI in the task of detection and segmentation of lesions suspicious for sPC. Materials and Methods In this retrospective study, T2-weighted and diffusion prostate MRI sequences from consecutive men examined with a single 3.0-T MRI system between 2015 and 2016 were manually segmented. Ground truth was provided by combined targeted and extended systematic MRI-transrectal US fusion biopsy, with sPC defined as International Society of Urological Pathology Gleason grade group greater than or equal to 2. By using split-sample validation, U-Net was internally validated on the training set (80% of the data) through cross validation and subsequently externally validated on the test set (20% of the data). U-Net-derived sPC probability maps were calibrated by matching sextant-based cross-validation performance to clinical performance of Prostate Imaging Reporting and Data System (PI-RADS). Performance of PI-RADS and U-Net were compared by using sensitivities, specificities, predictive values, and Dice coefficient. Results A total of 312 men (median age, 64 years; interquartile range [IQR], 58-71 years) were evaluated. The training set consisted of 250 men (median age, 64 years; IQR, 58-71 years) and the test set of 62 men (median age, 64 years; IQR, 60-69 years). In the test set, PI-RADS cutoffs greater than or equal to 3 versus cutoffs greater than or equal to 4 on a per-patient basis had sensitivity of 96% (25 of 26) versus 88% (23 of 26) at specificity of 22% (eight of 36) versus 50% (18 of 36). U-Net at probability thresholds of greater than or equal to 0.22 versus greater than or equal to 0.33 had sensitivity of 96% (25 of 26) versus 92% (24 of 26) (both P > .99) with specificity of 31% (11 of 36) versus 47% (17 of 36) (both P > .99), not statistically different from PI-RADS. Dice coefficients were 0.89 for prostate and 0.35 for MRI lesion segmentation. In the test set, coincidence of PI-RADS greater than or equal to 4 with U-Net lesions improved the positive predictive value from 48% (28 of 58) to 67% (24 of 36) for U-Net probability thresholds greater than or equal to 0.33 (P = .01), while the negative predictive value remained unchanged (83% [25 of 30] vs 83% [43 of 52]; P > .99). Conclusion U-Net trained with T2-weighted and diffusion MRI achieves similar performance to clinical Prostate Imaging Reporting and Data System assessment. © RSNA, 2019 Online supplemental material is available for this article. See also the editorial by Padhani and Turkbey in this issue.

Automated quantitative tumour response assessment of MRI in neuro-oncology with artificial neural networks: a multicentre, retrospective study

BACKGROUND

The Response Assessment in Neuro-Oncology (RANO) criteria and requirements for a uniform protocol have been introduced to standardise assessment of MRI scans in both clinical trials and clinical practice. However, these criteria mainly rely on manual two-dimensional measurements of contrast-enhancing (CE) target lesions and thus restrict both reliability and accurate assessment of tumour burden and treatment response. We aimed to develop a framework relying on artificial neural networks (ANNs) for fully automated quantitative analysis of MRI in neuro-oncology to overcome the inherent limitations of manual assessment of tumour burden.

METHODS

In this retrospective study, we compiled a single-institution dataset of MRI data from patients with brain tumours being treated at Heidelberg University Hospital (Heidelberg, Germany; Heidelberg training dataset) to develop and train an ANN for automated identification and volumetric segmentation of CE tumours and non-enhancing T2-signal abnormalities (NEs) on MRI. Independent testing and large-scale application of the ANN for tumour segmentation was done in a single-institution longitudinal testing dataset from the Heidelberg University Hospital and in a multi-institutional longitudinal testing dataset from the prospective randomised phase 2 and 3 European Organisation for Research and Treatment of Cancer (EORTC)-26101 trial (NCT01290939), acquired at 38 institutions across Europe. In both longitudinal datasets, spatial and temporal tumour volume dynamics were automatically quantified to calculate time to progression, which was compared with time to progression determined by RANO, both in terms of reliability and as a surrogate endpoint for predicting overall survival. We integrated this approach for fully automated quantitative analysis of MRI in neuro-oncology within an application-ready software infrastructure and applied it in a simulated clinical environment of patients with brain tumours from the Heidelberg University Hospital (Heidelberg simulation dataset).

FINDINGS

For training of the ANN, MRI data were collected from 455 patients with brain tumours (one MRI per patient) being treated at Heidelberg hospital between July 29, 2009, and March 17, 2017 (Heidelberg training dataset). For independent testing of the ANN, an independent longitudinal dataset of 40 patients, with data from 239 MRI scans, was collected at Heidelberg University Hospital in parallel with the training dataset (Heidelberg test dataset), and 2034 MRI scans from 532 patients at 34 institutions collected between Oct 26, 2011, and Dec 3, 2015, in the EORTC-26101 study were of sufficient quality to be included in the EORTC-26101 test dataset. The ANN yielded excellent performance for accurate detection and segmentation of CE tumours and NE volumes in both longitudinal test datasets (median DICE coefficient for CE tumours 0·89 [95% CI 0·86-0·90], and for NEs 0·93 [0·92-0·94] in the Heidelberg test dataset; CE tumours 0·91 [0·90-0·92], NEs 0·93 [0·93-0·94] in the EORTC-26101 test dataset). Time to progression from quantitative ANN-based assessment of tumour response was a significantly better surrogate endpoint than central RANO assessment for predicting overall survival in the EORTC-26101 test dataset (hazard ratios ANN 2·59 [95% CI 1·86-3·60] vs central RANO 2·07 [1·46-2·92]; p<0·0001) and also yielded a 36% margin over RANO (p<0·0001) when comparing reliability values (ie, agreement in the quantitative volumetrically defined time to progression [based on radiologist ground truth vs automated assessment with ANN] of 87% [266 of 306 with sufficient data] compared with 51% [155 of 306] with local vs independent central RANO assessment). In the Heidelberg simulation dataset, which comprised 466 patients with brain tumours, with 595 MRI scans obtained between April 27, and Sept 17, 2018, automated on-demand processing of MRI scans and quantitative tumour response assessment within the simulated clinical environment required 10 min of computation time (average per scan).

INTERPRETATION

Overall, we found that ANN enabled objective and automated assessment of tumour response in neuro-oncology at high throughput and could ultimately serve as a blueprint for the application of ANN in radiology to improve clinical decision making. Future research should focus on prospective validation within clinical trials and application for automated high-throughput imaging biomarker discovery and extension to other diseases.

FUNDING

Medical Faculty Heidelberg Postdoc-Program, Else Kröner-Fresenius Foundation.

The relevance of hippocampal subfield integrity and clock drawing test performance for the diagnosis of Alzheimer's disease and mild cognitive impairment

OBJECTIVES

The clock drawing test (CDT) is one of the worldwide most used screening tests for Alzheimer's disease (AD). MRI studies have identified temporo-parietal regions being involved in CDT impairment. However, the contributions of specific hippocampal subfields and adjacent extrahippocampal structures to CDT performance in AD and mild cognitive impairment (MCI) have not been investigated so far. It is unclear whether morphological alterations or CDT score, or a combination of both, are able to predict AD.

METHODS

38 AD patients, 38 MCI individuals and 31 healthy controls underwent neuropsychological assessment and MRI at 3 Tesla. FreeSurfer 5.3 was used to perform hippocampal parcellation. We used a collection of statistical methods to better understand the relationship between CDT and hippocampal formation. We also tested the clinical feasibility of this relationship when predicting AD.

RESULTS

Impaired CDT performance in AD was associated with widespread atrophy of the cornu ammonis, presubiculum, and subiculum, whereas MCI subjects showed CDT-related alterations of the CA4-dentate gyrus and subiculum. CDT correlates in AD and MCI showed regional and quantitative overlap. Importantly, CDT score was the best predictor of AD.

CONCLUSIONS

Our findings lend support for an involvement of different hippocampal subfields in impaired CDT performance in AD and MCI. CDT seems to be more efficient than subfield imaging for predicting AD.

Radiomic subtyping improves disease stratification beyond key molecular, clinical, and standard imaging characteristics in patients with glioblastoma

Background

The purpose of this study was to analyze the potential of radiomics for disease stratification beyond key molecular, clinical, and standard imaging features in patients with glioblastoma.

Methods

Quantitative imaging features (n = 1043) were extracted from the multiparametric MRI of 181 patients with newly diagnosed glioblastoma prior to standard-of-care treatment (allocated to a discovery and a validation set, 2:1 ratio). A subset of 386/1043 features were identified as reproducible (in an independent MRI test-retest cohort) and selected for analysis. A penalized Cox model with 10-fold cross-validation (Coxnet) was fitted on the discovery set to construct a radiomic signature for predicting progression-free and overall survival (PFS and OS). The incremental value of a radiomic signature beyond molecular (O6-methylguanine-DNA methyltransferase [MGMT] promoter methylation, DNA methylation subgroups), clinical (patient's age, KPS, extent of resection, adjuvant treatment), and standard imaging parameters (tumor volumes) for stratifying PFS and OS was assessed with multivariate Cox models (performance quantified with prediction error curves).

Results

The radiomic signature (constructed from 8/386 features identified through Coxnet) increased the prediction accuracy for PFS and OS (in both discovery and validation sets) beyond the assessed molecular, clinical, and standard imaging parameters (P ≤ 0.01). Prediction errors decreased by 36% for PFS and 37% for OS when adding the radiomic signature (compared with 29% and 27%, respectively, with molecular + clinical features alone). The radiomic signature was-along with MGMT status-the only parameter with independent significance on multivariate analysis (P ≤ 0.01).

Conclusions

Our study stresses the role of integrating radiomics into a multilayer decision framework with key molecular and clinical features to improve disease stratification and to potentially advance personalized treatment of patients with glioblastoma.

Automatic bone segmentation in whole-body CT images

PURPOSE

Many diagnostic or treatment planning applications critically depend on the successful localization of bony structures in CT images. Manual or semiautomatic bone segmentation is tedious, however, and often not practical in clinical routine. In this paper, we present a reliable and fully automatic bone segmentation in whole-body CT scans of patients suffering from multiple myeloma.

METHODS

We address this problem by using convolutional neural networks with an architecture inspired by the U-Net [17]. In this publication, we compared three training procedures: (1) training from 2D axial slices, (2) a pseudo-3D approach including axial, sagittal and coronal slices and (3) an approach where the network is pre-trained in an unsupervised manner.

RESULTS

We evaluated the method on an in-house dataset of 18 whole-body CT scans consisting of 6800 axial slices, achieving a dice score of 0.95 and an intersection over union (IOU) of 0.91. Furthermore, we evaluated our method on the dataset used by Peréz-Carrasco et al. (Comput Methods Progr Biomed 156:85-95, 2018). The data and the ground truth have been made publicly available. The proposed method outperformed the other methods, obtaining a dice score of 0.92 and an IOU of 0.85.

CONCLUSION

These promising results could facilitate the evaluation of bone density and the localization of focal lesions in the future, with a potential impact on both disease staging and treatment planning.

Deep Learning Techniques for Automatic MRI Cardiac Multi-Structures Segmentation and Diagnosis: Is the Problem Solved?

Delineation of the left ventricular cavity, myocardium, and right ventricle from cardiac magnetic resonance images (multi-slice 2-D cine MRI) is a common clinical task to establish diagnosis. The automation of the corresponding tasks has thus been the subject of intense research over the past decades. In this paper, we introduce the "Automatic Cardiac Diagnosis Challenge" dataset (ACDC), the largest publicly available and fully annotated dataset for the purpose of cardiac MRI (CMR) assessment. The dataset contains data from 150 multi-equipments CMRI recordings with reference measurements and classification from two medical experts. The overarching objective of this paper is to measure how far state-of-the-art deep learning methods can go at assessing CMRI, i.e., segmenting the myocardium and the two ventricles as well as classifying pathologies. In the wake of the 2017 MICCAI-ACDC challenge, we report results from deep learning methods provided by nine research groups for the segmentation task and four groups for the classification task. Results show that the best methods faithfully reproduce the expert analysis, leading to a mean value of 0.97 correlation score for the automatic extraction of clinical indices and an accuracy of 0.96 for automatic diagnosis. These results clearly open the door to highly accurate and fully automatic analysis of cardiac CMRI. We also identify scenarios for which deep learning methods are still failing. Both the dataset and detailed results are publicly available online, while the platform will remain open for new submissions.

TractSeg - Fast and accurate white matter tract segmentation

The individual course of white matter fiber tracts is an important factor for analysis of white matter characteristics in healthy and diseased brains. Diffusion-weighted MRI tractography in combination with region-based or clustering-based selection of streamlines is a unique combination of tools which enables the in-vivo delineation and analysis of anatomically well-known tracts. This, however, currently requires complex, computationally intensive processing pipelines which take a lot of time to set up. TractSeg is a novel convolutional neural network-based approach that directly segments tracts in the field of fiber orientation distribution function (fODF) peaks without using tractography, image registration or parcellation. We demonstrate that the proposed approach is much faster than existing methods while providing unprecedented accuracy, using a population of 105 subjects from the Human Connectome Project. We also show initial evidence that TractSeg is able to generalize to differently acquired data sets for most of the bundles. The code and data are openly available at https://github.com/MIC-DKFZ/TractSeg/ and https://doi.org/10.5281/zenodo.1088277, respectively.

Mask-Adapted Background Field Removal for Artifact Reduction in Quantitative Susceptibility Mapping of the Prostate

We propose an alternative processing method for quantitative susceptibility mapping of the prostate that reduces artifacts and enables better visibility and quantification of calcifications and other lesions. Three-dimensional gradient-echo magnetic resonance data were obtained from 26 patients at 3 T who previously received a planning computed tomography of the prostate. Phase images were unwrapped using Laplacian-based phase unwrapping. The background field was removed with the V-SHARP method using tissue masks for the entire abdomen (Method 1) and masks that excluded bone and the rectum (Method 2). Susceptibility maps were calculated with the iLSQR method. The quality of susceptibility maps was assessed by one radiologist and two physicists who rated the data for visibility of lesions and data quality on a scale from 1 (poor) to 4 (good). The readers rated susceptibility maps computed with Method 2 to be, on average, better for visibility of lesions with a score of 2.9 ± 1.1 and image quality with a score of 2.8 ± 0.8 compared with maps computed with Method 1 (2.4 ± 1.2/2.3 ± 1.0). Regarding strong artifacts, these could be removed using adapted masks, and the susceptibility values seemed less biased by the artifacts. Thus, using an adapted mask for background field removal when calculating susceptibility maps of the prostate from phase data reduces artifacts and improves visibility of lesions.

The challenge of mapping the human connectome based on diffusion tractography

Tractography based on non-invasive diffusion imaging is central to the study of human brain connectivity. To date, the approach has not been systematically validated in ground truth studies. Based on a simulated human brain data set with ground truth tracts, we organized an open international tractography challenge, which resulted in 96 distinct submissions from 20 research groups. Here, we report the encouraging finding that most state-of-the-art algorithms produce tractograms containing 90% of the ground truth bundles (to at least some extent). However, the same tractograms contain many more invalid than valid bundles, and half of these invalid bundles occur systematically across research groups. Taken together, our results demonstrate and confirm fundamental ambiguities inherent in tract reconstruction based on orientation information alone, which need to be considered when interpreting tractography and connectivity results. Our approach provides a novel framework for estimating reliability of tractography and encourages innovation to address its current limitations.

Though tractography is widely used, it has not been systematically validated. Here, authors report results from 20 groups showing that many tractography algorithms produce both valid and invalid bundles.

Effective user interaction in online interactive semantic segmentation of glioblastoma magnetic resonance imaging

Interactive segmentation is a promising approach to solving the pervasive shortage of reference annotations for automated medical image processing. We focus on the challenging task of glioblastoma segmentation in magnetic resonance imaging using a random forest pixel classifier trained iteratively on scribble annotations. Our experiments use data from the MICCAI Multimodal Brain Tumor Segmentation Challenge 2013 and simulate expert interactions using different approaches: corrective annotations, class-balanced corrections, annotations where classifier uncertainty is high, and corrections where classifier uncertainty is high/low. We find that it is better to correct the classifier than to provide annotations where the classifier is uncertain, resulting in significantly better Dice scores in the edema (0.662 to 0.686) and necrosis (0.550 to 0.676) regions after 20 interactions. It is also advantageous to balance inputs among classes, with significantly better Dice in the necrotic (0.501 to 0.676) and nonenhancing (0.151 to 0.235) regions compared to fully random corrections. Corrective annotations in regions of high classifier uncertainty provide no additional benefit, low uncertainty corrections perform worst. Preliminary experiments with real users indicate that those with intermediate proficiency make a considerable number of annotation errors. The performance of corrective approaches suffers most strongly from this, leading to a less profound difference to uncertainty-based annotations.

Fiber tractography using machine learning

We present a fiber tractography approach based on a random forest classification and voting process, guiding each step of the streamline progression by directly processing raw diffusion-weighted signal intensities. For comparison to the state-of-the-art, i.e. tractography pipelines that rely on mathematical modeling, we performed a quantitative and qualitative evaluation with multiple phantom and in vivo experiments, including a comparison to the 96 submissions of the ISMRM tractography challenge 2015. The results demonstrate the vast potential of machine learning for fiber tractography.

Cortical signature of clock drawing performance in Alzheimer's disease and mild cognitive impairment

It is unclear whether clock drawing test (CDT) performance relies on a widely distributed cortical network, or whether this test predominantly taps into parietal cortex function. So far, associations between cortical integrity and CDT impairment in Alzheimer's disease (AD) and mild cognitive impairment (MCI) largely stem from cortical volume analyses. Given that volume is a product of thickness and surface area, investigation of the relationship between CDT and these two cortical measures might contribute to better understanding of this cognitive screening tool for AD. 38 patients with AD, 38 individuals with MCI and 31 healthy controls (HC) underwent CDT assessment and MRI at 3 Tesla. The surface-based analysis via Freesurfer enabled calculation of cortical thickness and surface area. CDT was scored according to the method proposed by Shulman and related to the two distinct cortical measurements. Higher CDT scores across the entire sample were associated with cortical thickness in bilateral temporal gyrus, the right supramarginal gyrus, and the bilateral parietal gyrus, respectively (p < 0.001 CWP corr.). Significant associations between CDT and cortical thickness reduction in the parietal lobe remained significant when analyses were restricted to AD individuals. There was no statistically significant association between CDT scores and surface area (p < 0.001 CWP corr.). In conclusion, CDT performance may be driven by cortical thickness alterations in regions previously identified as "AD vulnerable", i.e. regions predominantly including temporal and parietal lobes. Our results suggest that cortical features of distinct evolutionary and genetic origin differently contribute to CDT performance.

White matter microstructure variations contribute to neurological soft signs in healthy adults

OBJECTIVE

Neurological soft signs (NSS) are core features of psychiatric disorders with significant neurodevelopmental origin. However, it is unclear whether NSS correlates are associated with neuropathological processes underlying the disease or if they are confounded by medication. Given that NSS are also present in healthy persons (HP), investigating HP could reveal NSS correlates, which are not biased by disease-specific processes or drug treatment. Therefore, we used a combination of diffusion MRI analysis tools to provide a framework of specific white matter (WM) microstructure variations underlying NSS in HP.

METHOD

NSS of 59 HP were examined on the Heidelberg Scale and related to diffusion associated metrics. Using tract-based spatial statistics (TBSS), we studied WM variations in fractional anisotropy (FA) as well as radial (RD), axial (AD), and mean diffusivity (MD). Using graph analytics (clustering coefficient-CC, local betweenness centrality -BC), we then explored DTI-derived structural network variations in regions identified by previous MRI studies on NSS.

RESULTS

NSS scores were negatively associated with RD, AD and MD in corpus callosum, brainstem and cerebellum (P < 0.05, corr.). NSS scores were negatively associated with CC and BC of the pallidum, the superior parietal gyrus, the precentral sulcus, the insula, and the cingulate gyrus (P < 0.05, uncorr.).

CONCLUSION

The present study supports the notion that WM microstructure variations in subcortical and cortical sensorimotor regions contribute to NSS expression in young HP. Hum Brain Mapp 38:3552-3565, 2017. © 2017 Wiley Periodicals, Inc.

Hippocampal formation alterations differently contribute to autobiographic memory deficits in mild cognitive impairment and Alzheimer's disease

Autobiographical memory (AM) is part of declarative memory and includes both semantic and episodic aspects. AM deficits are among the major complaints of patients with Alzheimer's disease (AD) even in early or preclinical stages. Previous MRI studies in AD patients have showed that deficits in semantic and episodic AM are associated with hippocampal alterations. However, the question which specific hippocampal subfields and adjacent extrahippocampal structures contribute to deficits of AM in individuals with mild cognitive impairment (MCI) and AD patients has not been investigated so far. Hundred and seven participants (38 AD patients, 38 MCI individuals and 31 healthy controls [HC]) underwent MRI at 3 Tesla. AM was assessed with a semi-structured interview (E-AGI). FreeSurfer 5.3 was used for hippocampal parcellation. Semantic and episodic AM scores were related to the volume of 5 hippocampal subfields and cortical thickness in the parahippocampal and entorhinal cortex. Both semantic and episodic AM deficits were associated with bilateral hippocampal alterations. These associations referred mainly to CA1, CA2-3, presubiculum, and subiculum atrophy. Episodic, but not semantic AM loss was associated with cortical thickness reduction of the bilateral parahippocampal and enthorinal cortex. In MCI individuals, episodic, but not semantic AM deficits were associated with alterations of the CA1, presubiculum and subiculum. Our findings support the crucial role of CA1, presubiculum, and subiculum in episodic memory. The present results implicate that in MCI individuals, semantic and episodic AM deficits are subserved by distinct neuronal systems.

Prediction of malignancy by a radiomic signature from contrast agent-free diffusion MRI in suspicious breast lesions found on screening mammography

PURPOSE

To assess radiomics as a tool to determine how well lesions found suspicious on breast cancer screening X-ray mammography can be categorized into malignant and benign with unenhanced magnetic resonance (MR) mammography with diffusion-weighted imaging and T₂ -weighted sequences.

MATERIALS AND METHODS

From an asymptomatic screening cohort, 50 women with mammographically suspicious findings were examined with contrast-enhanced breast MRI (ceMRI) at 1.5T. Out of this protocol an unenhanced, abbreviated diffusion-weighted imaging protocol (ueMRI) including T₂ -weighted, (T₂ w), diffusion-weighted imaging (DWI), and DWI with background suppression (DWIBS) sequences and corresponding apparent diffusion coefficient (ADC) maps were extracted. From ueMRI-derived radiomic features, three Lasso-supervised machine-learning classifiers were constructed and compared with the clinical performance of a highly experienced radiologist: 1) univariate mean ADC model, 2) unconstrained radiomic model, 3) constrained radiomic model with mandatory inclusion of mean ADC.

RESULTS

The unconstrained and constrained radiomic classifiers consisted of 11 parameters each and achieved differentiation of malignant from benign lesions with a .632 + bootstrap receiver operating characteristics (ROC) area under the curve (AUC) of 84.2%/85.1%, compared to 77.4% for mean ADC and 95.9%/95.9% for the experienced radiologist using ceMRI/ueMRI.

CONCLUSION

In this pilot study we identified two ueMRI radiomics classifiers that performed well in the differentiation of malignant from benign lesions and achieved higher performance than the mean ADC parameter alone. Classification was lower than the almost perfect performance of a highly experienced breast radiologist. The potential of radiomics to provide a training-independent diagnostic decision tool is indicated. A performance reaching the human expert would be highly desirable and based on our results is considered possible when the concept is extended in larger cohorts with further development and validation of the technique.

LEVEL OF EVIDENCE

1 Technical Efficacy: Stage 2 J. MAGN. RESON. IMAGING 2017;46:604-616.