Abstract
Clinical assessment of moderate-to-severe traumatic brain injury (msTBI) relies heavily on manual review of structural MRI by trained readers. Automated segmentation of msTBI lesions on T1-weighted MRI aims to relieve some of this burden. However, deep learning-based approaches remain challenging due to the extreme variability in lesion size, shape, and location. In this work, we describe our entry to the AIMS-TBI 2026 Challenge. We adapt a 3D U-Net (~40M parameters) with residual connections and deep supervision initialized from a model trained on the ATLAS v3.0 dataset. The model was trained with a composite loss combining binary cross-entropy, dice loss, and a lesion-wise dice term on both exams with positive findings and no findings. Our submission achieved a balanced accuracy of 0.871 and a dice similarity coefficient of 0.531 on the held-out test set.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AIMS_TBI_010.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=~Jon_Andr%C3%A9_Ottesen1
BibTex
@InProceedings{OttJon_AUNet_MICCAISAT2026,
author = { Ottesen, Jon André AND Liu, Qinghui AND MacIntosh, Bradley AND Emblem, Kyrre Eeg AND Bjørnerud, Atle},
title = { { A U-Net for Automated msTBI Lesion 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}
}
