Abstract
Accurate lesion segmentation in moderate-to-severe traumatic brain injury (ms-TBI) from T1-weighted (T1-w) MRI remains challenging because lesions vary considerably in appearance, location, number, and size. As T1-w MRI is widely available in clinical and research settings, an automatic method based on this sequence alone could support the analysis of large-scale ms-TBI datasets. We present our approach for the Automated Identification of Moderate-to-Severe TBI Lesions (AIMS-TBI 2026) challenge. An nnU-Net-based architecture was initialized from MultiTalentV2 weights obtained through cross-modality supervised pretraining on heterogeneous medical datasets. A hierarchical multi-patch schedule adapted one model from larger contextual patches to smaller lesion-centred patches, while a second model was fine-tuned independently. The final pipeline combined two complementary models using probability averaging and connected component filtering. On the Phase-2 validation dataset, the submitted ensemble achieved a lesion-positive Dice of 0.609; across all scans, it obtained a Dice of 0.723, which is influenced by lesion-free scans. On the hidden test set, our method achieved a lesion-positive Dice of 0.5533, HD95 of 28.29 mm, and ASSD of 6.67 mm, ranking third on the challenge leaderboard. These results show the feasibility of a T1-w-only ms-TBI lesion-segmentation pipeline using cross-modality supervised initialization, targeted fine-tuning, and probability-map ensembling.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AIMS_TBI_007.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Harshit_Shukla2
BibTex
@InProceedings{ShuHar_Hybrid_MICCAISAT2026,
author = { Shukla, Harshit AND Masilamani, V. G. AND Mohanraj, Aravindhan AND Krishnamurthi, Ganapathy},
title = { { Hybrid Ensemble with Multi-Patch Fine-Tuning for Traumatic Brain Injury Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17254},
month = {pending},
page = {pending}
}