LST-AI: a Deep Learning Ensemble for Accurate MS Lesion Segmentation

Automated segmentation of brain white matter lesions is crucial for both clinical assessment and scientific research in multiple sclerosis (MS). Over a decade ago, we introduced an engineered lesion segmentation tool, LST. While recent lesion segmentation approaches have leveraged artificial intelligence (AI), they often remain proprietary and difficult to adopt. As an open-source tool, we present LST-AI, an advanced deep learning-based extension of LST that consists of an ensemble of three 3D-UNets. LST-AI explicitly addresses the imbalance between white matter (WM) lesions and non-lesioned WM. It employs a composite loss function incorporating binary cross-entropy and Tversky loss to improve segmentation of the highly heterogeneous MS lesions. We train the network ensemble on 491 MS pairs of T1w and FLAIR images, collected in-house from a 3T MRI scanner, and expert neuroradiologists manually segmented the utilized lesion maps for training. LST-AI additionally includes a lesion location annotation tool, labeling lesion location according to the 2017 McDonald criteria (periventricular, infratentorial, juxtacortical, subcortical). We conduct evaluations on 103 test cases consisting of publicly available data using the Anima segmentation validation tools and compare LST-AI with several publicly available lesion segmentation models. Our empirical analysis shows that LST-AI achieves superior performance compared to existing methods. Its Dice and F1 scores exceeded 0.62, outperforming LST, SAMSEG (Sequence Adaptive Multimodal SEGmentation), and the popular nnUNet framework, which all scored below 0.56. Notably, LST-AI demonstrated exceptional performance on the MSSEG-1 challenge dataset, an international WM lesion segmentation challenge, with a Dice score of 0.65 and an F1 score of 0.63-surpassing all other competing models at the time of the challenge. With increasing lesion volume, the lesion detection rate rapidly increased with a detection rate of >75% for lesions with a volume between 10mm3 and 100mm3. Given its higher segmentation performance, we recommend that research groups currently using LST transition to LST-AI. To facilitate broad adoption, we are releasing LST-AI as an open-source model, available as a command-line tool, dockerized container, or Python script, enabling diverse applications across multiple platforms.

Retrospective cohort study to devise a treatment decision score predicting adverse 24-month radiological activity in early multiple sclerosis

Background

Multiple sclerosis (MS) is a chronic neuroinflammatory disease affecting about 2.8 million people worldwide. Disease course after the most common diagnoses of relapsing-remitting multiple sclerosis (RRMS) and clinically isolated syndrome (CIS) is highly variable and cannot be reliably predicted. This impairs early personalized treatment decisions.

Objectives

The main objective of this study was to algorithmically support clinical decision-making regarding the options of early platform medication or no immediate treatment of patients with early RRMS and CIS.

Design

Retrospective monocentric cohort study within the Data Integration for Future Medicine (DIFUTURE) Consortium.

Methods

Multiple data sources of routine clinical, imaging and laboratory data derived from a large and deeply characterized cohort of patients with MS were integrated to conduct a retrospective study to create and internally validate a treatment decision score [Multiple Sclerosis Treatment Decision Score (MS-TDS)] through model-based random forests (RFs). The MS-TDS predicts the probability of no new or enlarging lesions in cerebral magnetic resonance images (cMRIs) between 6 and 24 months after the first cMRI.

Results

Data from 65 predictors collected for 475 patients between 2008 and 2017 were included. No medication and platform medication were administered to 277 (58.3%) and 198 (41.7%) patients. The MS-TDS predicted individual outcomes with a cross-validated area under the receiver operating characteristics curve (AUROC) of 0.624. The respective RF prediction model provides patient-specific MS-TDS and probabilities of treatment success. The latter may increase by 5-20% for half of the patients if the treatment considered superior by the MS-TDS is used.

Conclusion

Routine clinical data from multiple sources can be successfully integrated to build prediction models to support treatment decision-making. In this study, the resulting MS-TDS estimates individualized treatment success probabilities that can identify patients who benefit from early platform medication. External validation of the MS-TDS is required, and a prospective study is currently being conducted. In addition, the clinical relevance of the MS-TDS needs to be established.

Lesion location across diagnostic regions in multiple sclerosis

BACKGROUND

Lesions in the periventricular, (juxta)cortical, and infratentorial region, as visible on brain MRI, are part of the diagnostic criteria for Multiple sclerosis (MS) whereas lesions in the subcortical region are currently only a marker of disease activity. It is unknown whether MS lesions follow individual spatial patterns or whether they occur in a random manner across diagnostic regions.

AIM

First, to describe cross-sectionally the spatial lesion patterns in patients with MS. Second, to investigate the spatial association of new lesions and lesions at baseline across diagnostic regions.

METHODS

Experienced neuroradiologists analyzed brain MRI (3D, 3T) in a cohort of 330 early MS patients. Lesions at baseline and new solitary lesions after two years were segmented (manually and by consensus) and classified as periventricular, (juxta)cortical, or infratentorial (diagnostic regions) or subcortical-with or without Gadolinium-enhancement. Gadolinium enhancement of lesions in the different regions was compared by Chi square test. New lesions in the four regions served as dependent variable in four zero-inflated Poisson models each with the six independent variables of lesions in the four regions at baseline, age and gender.

RESULTS

At baseline, lesions were most often observed in the subcortical region (mean 13.0 lesions/patient), while lesion volume was highest in the periventricular region (mean 2287 µl/patient). Subcortical lesions were less likely to show gadolinium enhancement (3.1 %) than juxtacortical (4.3 %), periventricular (5.3 %) or infratentorial lesions (7.2 %). Age was inversely correlated with new periventricular, juxtacortical and subcortical lesions. New lesions in the periventricular, juxtacortical and infratentorial region showed a significant autocorrelative behavior being positively related to the number of lesions in the respective regions at baseline. New lesions in the subcortical region showed a different behavior with a positive association with baseline periventricular lesions and a negative association with baseline infratentorial lesions.

CONCLUSION

Across regions, new lesions do not occur randomly; instead, new lesions in the periventricular, juxtacortical and infratentorial diagnostic region are associated with that at baseline. Lesions in the subcortical regions are more closely related to periventricular lesions. Moreover, subcortical lesions substantially contribute to lesion burden in MS but are less likely to show gadolinium enhancement (than lesions in the diagnostic regions).

Uncertainty-Aware and Lesion-Specific Image Synthesis in Multiple Sclerosis Magnetic Resonance Imaging: A Multicentric Validation Study

Generative adversarial networks (GANs) can synthesize high-contrast MRI from lower-contrast input. Targeted translation of parenchymal lesions in multiple sclerosis (MS), as well as visualization of model confidence further augment their utility, provided that the GAN generalizes reliably across different scanners. We here investigate the generalizability of a refined GAN for synthesizing high-contrast double inversion recovery (DIR) images and propose the use of uncertainty maps to further enhance its clinical utility and trustworthiness. A GAN was trained to synthesize DIR from input fluid-attenuated inversion recovery (FLAIR) and T1w of 50 MS patients (training data). In another 50 patients (test data), two blinded readers (R1 and R2) independently quantified lesions in synthetic DIR (synthDIR), acquired DIR (trueDIR) and FLAIR. Of the 50 test patients, 20 were acquired on the same scanner as training data (internal data), while 30 were scanned at different scanners with heterogeneous field strengths and protocols (external data). Lesion-to-Background ratios (LBR) for MS-lesions vs. normal appearing white matter, as well as image quality parameters were calculated. Uncertainty maps were generated to visualize model confidence. Significantly more MS-specific lesions were found in synthDIR compared to FLAIR (R1: 26.7 ± 2.6 vs. 22.5 ± 2.2 p < 0.0001; R2: 22.8 ± 2.2 vs. 19.9 ± 2.0, p = 0.0005). While trueDIR remained superior to synthDIR in R1 [28.6 ± 2.9 vs. 26.7 ± 2.6 (p = 0.0021)], both sequences showed comparable lesion conspicuity in R2 [23.3 ± 2.4 vs. 22.8 ± 2.2 (p = 0.98)]. Importantly, improvements in lesion counts were similar in internal and external data. Measurements of LBR confirmed that lesion-focused GAN training significantly improved lesion conspicuity. The use of uncertainty maps furthermore helped discriminate between MS lesions and artifacts. In conclusion, this multicentric study confirms the external validity of a lesion-focused Deep-Learning tool aimed at MS imaging. When implemented, uncertainty maps are promising to increase the trustworthiness of synthetic MRI.

Improving Automated Glioma Segmentation in Routine Clinical Use Through Artificial Intelligence-Based Replacement of Missing Sequences With Synthetic Magnetic Resonance Imaging Scans

OBJECTIVES

Although automated glioma segmentation holds promise for objective assessment of tumor biology and response, its routine clinical use is impaired by missing sequences, for example, due to motion artifacts. The aim of our study was to develop and validate a generative adversarial network for synthesizing missing sequences to allow for a robust automated segmentation.

MATERIALS AND METHODS

Our model was trained on data from The Cancer Imaging Archive (n = 238 WHO II-IV gliomas) to synthesize either missing FLAIR, T2-weighted, T1-weighted (T1w), or contrast-enhanced T1w images from available sequences, using a novel tumor-targeting loss to improve synthesis of tumor areas. We validated performance in a test set from both the REMBRANDT repository and our local institution (n = 68 WHO II-IV gliomas), using qualitative image appearance metrics, but also segmentation performance with state-of-the-art segmentation models. Segmentation of synthetic images was compared with 2 commonly used strategies for handling missing input data, entering a blank mask or copying an existing sequence.

RESULTS

Across tumor areas and missing sequences, synthetic images generally outperformed both conventional approaches, in particular when FLAIR was missing. Here, for edema and whole tumor segmentation, we improved the Dice score, a common metric for evaluation of segmentation performance, by 12% and 11%, respectively, over the best conventional method. No method was able to reliably replace missing contrast-enhanced T1w images.

DISCUSSION

Replacing missing nonenhanced magnetic resonance sequences via synthetic images significantly improves segmentation quality over most conventional approaches. This model is freely available and facilitates more widespread use of automated segmentation in routine clinical use, where missing sequences are common.

Automated Pathology Detection and Patient Triage in Routinely Acquired Head Computed Tomography Scans

OBJECTIVES

Anomaly detection systems can potentially uncover the entire spectrum of pathologies through deviations from a learned norm, meaningfully supporting the radiologist's workflow. We aim to report on the utility of a weakly supervised machine learning (ML) tool to detect pathologies in head computed tomography (CT) and adequately triage patients in an unselected patient cohort.

MATERIALS AND METHODS

All patients having undergone a head CT at a tertiary care hospital in March 2020 were eligible for retrospective analysis. Only the first scan of each patient was included. Anomaly detection was performed using a weakly supervised ML technique. Anomalous findings were displayed on voxel-level and pooled to an anomaly score ranging from 0 to 1. Thresholds for this score classified patients into the 3 classes: "normal," "pathological," or "inconclusive." Expert-validated radiological reports with multiclass pathology labels were considered as ground truth. Test assessment was performed with receiver operator characteristics analysis; inconclusive results were pooled to "pathological" predictions for accuracy measurements. External validity was tested in a publicly available external data set (CQ500).

RESULTS

During the investigation period, 297 patients were referred for head CT of which 248 could be included. Definite ratings into normal/pathological were feasible in 167 patients (67.3%); 81 scans (32.7%) remained inconclusive. The area under the curve to differentiate normal from pathological scans was 0.95 (95% confidence interval, 0.92-0.98) for the study data set and 0.87 (95% confidence interval, 0.81-0.94) in external validation. The negative predictive value to exclude pathology if a scan was classified as "normal" was 100% (25/25), and the positive predictive value was 97.6% (137/141). Sensitivity and specificity were 100% and 86%, respectively. In patients with inconclusive ratings, pathologies were found in 26 (63%) of 41 cases.

CONCLUSIONS

Our study provides the first clinical evaluation of a weakly supervised anomaly detection system for brain imaging. In an unselected, consecutive patient cohort, definite classification into normal/diseased was feasible in approximately two thirds of scans, going along with an excellent diagnostic accuracy and perfect negative predictive value for excluding pathology. Moreover, anomaly heat maps provide important guidance toward pathology interpretation, also in cases with inconclusive ratings.

Automated Detection of Ischemic Stroke and Subsequent Patient Triage in Routinely Acquired Head CT

Purpose

Advanced machine-learning (ML) techniques can potentially detect the entire spectrum of pathology through deviations from a learned norm.

We investigated the utility of a weakly supervised ML tool to detect characteristic findings related to ischemic stroke in head CT and provide subsequent patient triage.

Methods

Patients having undergone non-enhanced head CT at a tertiary care hospital in April 2020 with either no anomalies, subacute or chronic ischemia, lacunar infarcts of the deep white matter or hyperdense vessel signs were retrospectively analyzed. Anomaly detection was performed using a weakly supervised ML classifier. Findings were displayed on a voxel-level (heatmap) and pooled to an anomaly score. Thresholds for this score classified patients into i) normal, ii) inconclusive, iii) pathological. Expert-validated radiological reports were considered as ground truth. Test assessment was performed with ROC analysis; inconclusive results were pooled to pathological predictions for accuracy measurements.

Results

During the investigation period 208 patients were referred for head CT of which 111 could be included. Definite ratings into normal/pathological were feasible in 77 (69.4%) patients. Based on anomaly scores, the AUC to differentiate normal from pathological scans was 0.98 (95% CI 0.97–1.00). The sensitivity, specificity, positive and negative predictive values were 100%, 40.6%, 80.6% and 100%, respectively.

Conclusion

Our study demonstrates the potential of a weakly supervised anomaly-detection tool to detect stroke findings in head CT. Definite classification into normal/pathological was made with high accuracy in > 2/3 of patients. Anomaly heatmaps further provide guidance towards pathologies, also in cases with inconclusive ratings.

Local Bone Mineral Density, Subcutaneous and Visceral Adipose Tissue Measurements in Routine Multi Detector Computed Tomography-Which Parameter Predicts Incident Vertebral Fractures Best?

In this case-control study the value of bone mineral density (BMD) at different vertebral levels, subcutaneous adipose tissue (SAT), and visceral adipose tissue (VAT) to identify patients with incident osteoporotic vertebral fractures in routine multi-detector computed tomography (MDCT) exams was assessed.

MATERIAL AND METHODS

Seventeen patients who underwent baseline and follow-up routine contrast-enhanced MDCT and had an incident osteoporotic vertebral fracture at follow-up were included. Seventeen age-, sex- and follow-up duration-matched controls were identified. Trabecular BMD (from Th5 to L5) as well as cross-sectional area of SAT and VAT were extracted.

RESULTS

BMD performed best to differentiate patients with an incident fracture from controls at the levels of Th5 (area under the curve [AUC] = 0.781, p = 0.014), Th7 (AUC = 0.877, p = 0.001), and Th9 (AUC = 0.818, p = 0.005). Applying multivariate logistic regression BMD at Th7 level remained the only significant predictor of incident vertebral fractures (Th5-L5) with an odds ratio of 1.07 per BMD SD decrease. VAT and SAT did not show significant differences between the fracture and control group (p > 0.05).

CONCLUSION

The local BMD measurement appears to be more suitable than standard mean BMD from L1-L3 for fracture risk assessment.

Association of quadriceps muscle, gluteal muscle, and femoral bone marrow composition using chemical shift encoding-based water-fat MRI: a preliminary study in healthy young volunteers

Background

We investigated the composition of the gluteal (gluteus maximus, medius, and minimus) and quadriceps (rectus femoris, vastus lateralis, medialis, and intermedius) muscle groups and its associations with femoral bone marrow using chemical shift encoding-based water-fat magnetic resonance imaging (CSE-MRI) to improve our understanding of muscle-bone interaction.

Methods

Thirty healthy volunteers (15 males, aged 30.5 ± 4.9 years [mean ± standard deviation]; 15 females, aged 29.9 ± 7.1 years) were recruited. A six-echo three-dimensional spoiled gradient-echo sequence was used for 3-T CSE-MRI at the thigh and hip region. The proton density fat fraction (PDFF) of the gluteal and quadriceps muscle groups as well as of the femoral head, neck, and greater trochanter bone marrow were extracted and averaged over both sides.

Results

PDFF values of all analysed bone marrow compartments were significantly higher in men than in women (p ≤ 0.047). PDFF values of the analysed muscles showed no significant difference between men and women (p ≥ 0.707). After adjusting for age and body mass index, moderate significant correlations of PDFF values were observed between the gluteal and quadriceps muscle groups (r = 0.670) and between femoral subregions (from r = 0.613 to r = 0.655). Regarding muscle-bone interactions, only the PDFF of the quadriceps muscle and greater trochanter bone marrow showed a significant correlation (r = 0.375).

Conclusions

The composition of the muscle and bone marrow compartments at the thigh and hip region in young, healthy subjects seems to be quite distinct, without evidence for a strong muscle-bone interaction.