Abstract
Automated lesion detection and segmentation in moderate-to-severe traumatic brain injury (msTBI) remains challenging due to the heterogeneous appearance of lesions. The AIMS-TBI 2026 Challenge evaluates T1-weighted MRI lesion detection and segmentation while permitting additional data and modalities during training. We investigate three complementary strategies for improving T1-only performance: initialization from supervised MultiTalent and self-supervised OpenMind pretrained checkpoints, privileged diffusion MRI supervision through auxiliary FA/MD regression, and fine-tuning on additional annotated challenge data. Although both pretraining approaches improved the ResEncL architecture over random initialization, neither matched the PlainConv nnU-Net baseline, and auxiliary diffusion supervision provided no measurable benefit. Error analysis showed that this performance gap was driven primarily by complete lesion misses rather than poorer delineation of detected lesions. Different models exhibited complementary detections, motivating a hard-mask union of the PlainConv and MultiTalent-initialized ResEncL models, which outperformed the individual models as well as probability averaging and majority voting. Fine-tuning the PlainConv model on additional validation cases further improved segmentation, while optimizing the detection operating point of the PlainConv-ResEncL union improved balanced accuracy. Our results show that, for T1-only msTBI lesion analysis, identifying model failure modes and exploiting complementary detections can be more effective than increasing model complexity or pretraining scale.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AIMS_TBI_009.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=~Katharina_Eckstein1
BibTex
@InProceedings{EckKat_Evaluating_MICCAISAT2026,
author = { Eckstein, Katharina AND Kovacs, Balint AND Ulrich, Constantin AND Peretzke, Robin AND Maier-Hein, Klaus H.},
title = { { Evaluating Large-Scale Pretraining and Auxiliary Training Data for T1-only TBI Lesion Detection and Segmentation in the AIMS-TBI 2026 Challenge } },
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}
}
