Segmentation multimodale et multiclasse des lésions de la moelle épinière par apprentissage profond

Abstract

This study addresses automatic segmentation of the two pathological components of traumatic spinal cord lesions—edema and hemorrhage—from multimodal MRI, and their quantification for prognostic assessment. The database comprises 148 patients recruited at the Centre de recherche de l’Hôpital du Sacré-Coeur de Montréal, with manual annotations on 502 sagittal slices (140 with T1/T2 sagittal, 122 with axial T2). A pipeline integrates preprocessing, cascade segmentation, vertebral-level biomarkers and explainability. Two main contributions were developed : a supervised multiclass cascade that significantly outperforms the state of the art on this cohort and distinguishes edema from hemorrhage, and a prognostic analysis showing that differentiated biomarkers add predictive value for medium- and long-term neurological outcome beyond total lesion volume and clinical baseline. The first contribution is a cascade of two 2D U-Nets exploiting complementary sagittal and axial information. The sagittal model (T1, T2, STIR, cord mask) produces an initial segmentation ; its predictions are resampled to axial space as an extra input channel for the second model, with a topology loss enforcing hemorrhage surrounded by edema. Compared to SciSeg (Spinal Cord Toolbox), our method achieves median Dice improvement of 0.250 in the axial plane (0.540 vs 0.268, P < 0.001) and 0.054 in the sagittal plane (0.575 vs 0.461, P < 0.001), with 5-fold cross-validation. Edema segmentation is satisfactory (Dice 0.557) ; hemorrhage is limited (Dice 0.135) due to class imbalance and rarity. The second contribution quantifies prognostic biomarkers per vertebral level (area, volume, hemorrhage-to-lesion ratio) and evaluates their predictive value for ASIA improvement. Nested linear models—clinical baseline (M1), plus total lesion volume (M2), plus differentiated biomarkers (M3)—were compared at several follow-up times in 39 patients with longitudinal data. From 3 months post-surgery onward, differentiated biomarkers significantly improved prediction of neurological recovery compared to total volume alone (P = 0.028), with even stronger effects at 6 and 12 months (P < 0.001), while no gain was observed at the immediate post-operative assessment (P = 0.258). SHAP analysis indicated contextual features for hemorrhage detection ; a visualization prototype links metrics to axial MRI and segmentations.

Date	27 May 2026
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Luc Duong (Supervisor), Sylvie Ratté (Co-supervisor) & Jean Marc Mac-Thiong (Co-supervisor)