Self-supervised Pretraining and Multi-label Decoding for Intracranial Hemorrhage Segmentation

Aloys Portafaix, Laurent Létourneau-Guillon, Samuel Kadoury

Communication écrite (2025)

Accéder à ce document
Lien externe
Ce document n'est pas archivé dans PolyPublie, mais peut être consulté chez l'éditeur officiel

Résumé

Automated segmentation of intracranial hemorrhage (ICH) is an important clinical task limited by the scarcity of high-quality annotated data and variability across datasets. To address these challenges, we propose a novel two-phase learning framework that leverages self-supervised pretraining and multi-label decoding. First, we pretrain a hierarchical vision transformer on over 25,000 head CT cases using a Masked Autoencoder (MAE) strategy to learn robust multi-scale feature representations. For the downstream task, we append a multi-label decoder with dataset-specific heads, which enables joint learning from four distinct ICH segmentation datasets with diverse annotation protocols. A tailored sampling strategy further addresses class imbalance and heterogeneity across datasets. Our results demonstrate superior performance, with the proposed model achieving a mean Dice score of 63.57% (±12.30), surpassing the next best-performing method by 2.69%. Notably, the proposed MAE-based model outperformed its BYOL-pretrained counterpart, suggesting that MAE’s reconstruction-based objective is a more effective pretraining strategy for learning the dense, spatially-rich representations required for this segmentation task.

Mots clés

Renseignements supplémentaires:
5th International Workshop on Stroke Imaging and Treatment, SWITCH 2025, held in conjunction with MICCAI 2025, in Daejeon, South Korea, on September 23, 2025. ;
TopBrain Segmentation Challenge for Whole Brain Vessel Anatomy and SWITCH2025: Stroke Workshop on Imaging and Treatment CHallenges
Organismes subventionnaires:
Quebec Bio-Imaging Network, FRQ-S, FARQ
Numéro de subvention:
35450
ISBN:
9783032079459
Adresse URL de PolyPublie:
Nom de la conférence:
5th International Workshop on Stroke Imaging and Treatment (SWITCH 2025)
Lieu de la conférence:
Daejeon, South Korea
Date(s) de la conférence:
2025-09-23
Maison d'édition:
Springer Nature
OAI:
oai:publications.polymtl.ca:69063
ORCID
Date du dépôt:
10 oct. 2025 11:31
Dernière modification:
10 nov. 2026 10:19
Citer en APA 7:
Portafaix, A., Létourneau-Guillon, L., & Kadoury, S. (septembre 2025). Self-supervised Pretraining and Multi-label Decoding for Intracranial Hemorrhage Segmentation [Communication écrite]. 5th International Workshop on Stroke Imaging and Treatment (SWITCH 2025), Daejeon, South Korea. https://doi.org/10.1007/978-3-032-07945-9_9

Statistiques

Dimensions

Actions réservées au personnel

Afficher document
Afficher document