Robust automated detection of central vein sign and paramagnetic rim lesions in multiple sclerosis: An external validation study

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électricien, à finalité spécialisée

Abstract

enMultiple sclerosis (MS) is a chronic autoimmune disease of the central nervous system that causes progressive neurological disability. Magnetic resonance imaging (MRI) plays a central role in its diagnosis and monitoring, as it allows the detection and characterisation of white matter lesions in the brain. However, lesion-level analysis currently relies on manual annotation by trained raters, which is time-consuming and susceptible to inter-rater variability. This motivates the development of automated methods capable of extracting lesion-level features directly from MRI. Among the most informative lesion-level biomarkers are the central vein sign (CVS) and paramagnetic rim lesions (PRL). Although both are specific to MS, they provide distinct diagnostic insights. CVS reflects the perivenular origin of MS lesions and helps distinguish MS from other white matter diseases, while PRL marks lesions with ongoing chronic active inflammation, which is associated with greater disability progression. Their inclusion in the 2024 revision of the McDonald diagnostic criteria marks a shift from research tools to clinically recognised diagnostic markers. A particular challenge for automated methods is the presence of confluent lesions, which arise when distinct MS lesions merge into a single connected region. This complicates lesion counting, boundary delineation, and the assignment of lesion-level biomarkers such as CVS and PRL. This thesis evaluates Automated Lesion, PRL and CVS Analysis (ALPaCA), an automated deep learning method that simultaneously segments MS lesions and predicts CVS and PRL status from multi-contrast MRI. ALPaCA is applied to an external dataset of 63 MS patients, distinct from the one used in the original ALPaCA study, to assess its generalisation capacity. Performance is compared against the original paper results and against two baseline methods: Automated PRL (APRL) for PRL detection, and FLAIR Lesion Analysis in Multiple Sclerosis (FLAMeS) as an alternative segmentation backbone replacing Method for Inter-Modal Segmentation Analysis (MIMoSA) within the ALPaCA pipeline. Complementary evaluation metrics, including instance-level detection metrics and metrics restricted to confluent lesion units, are computed to provide a more complete picture of model performance. The results show a substantial drop in performance on the external dataset, with area under the receiver operating characteristic curve (AUROC) falling from 0.95 to 0.62 for lesion detection and from 0.91 to 0.62 for PRL detection. The main source of failure is identified at the candidate generation stage: the permissive MIMoSA threshold of 0.05 produces many false positive candidates that the downstream convolutional neural network (CNN) fails to filter out on out-of-distribution data. Replacing MIMoSA with FLAMeS improves lesion detection across all standard metrics. For PRL detection, no method clearly dominates the other on the full cohort; meaningful differences only emerge when restricting the analysis to subjects with at least one confirmed PRL, which points to the need for subset analyses when biomarkers are sparse. CVS evaluation could not be completed due to annotation inconsistencies in the reference masks, and a structured evaluation procedure is proposed for use once a corrected reference mask becomes available.