Abstract
Moderate-to-severe traumatic brain injury (msTBI) produces lesions that are focal or diffuse, span grey matter, white matter and cerebrospinal fluid, and vary over four orders of magnitude in volume. The AIMS-TBI 2026 challenge scores lesion detection and lesion segmentation on two separate leaderboards, on T1-weighted MRI only. We submitted a single self-configuring 3D residual-encoder U-Net that serves both tasks: the segmentation mask is the network output, and the detection decision is derived from it by connected component analysis at the challenge’s own 10-voxel floor. Three design choices were settled experimentally. First, we enlarged the training cohort by merging the 2026 training release with the labelled 2025 validation release, giving 652 cases. Second, on cases no constituent model had seen, no five-fold ensemble beat the best single fold, so we shipped the cheaper single fold, though the margin is schedule-dependent. Third, removing predicted components below the challenge’s 50-voxel scoring floor raised specificity but degraded lesion-wise detection rate, Dice and count error, so predictions are submitted unfiltered. On a 131-case internal held-out split the model reaches a mean Dice of 0.577 on lesion-containing scans and a balanced accuracy of 0.746; on the official Test Phase 1 data it reaches a Dice of 0.5154 and a detection score of 0.8331 (sensitivity 0.7188, specificity 0.9474). Residual error is concentrated in small lesions: mean Dice falls steeply with ground-truth lesion volume, and below roughly 500 voxels the network frequently predicts an empty mask.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AIMS_TBI_011.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=~Ujjwal_Baid2
BibTex
@InProceedings{BaiUjj_SingleModel_MICCAISAT2026,
author = { Baid, Ujjwal AND Baheti, Bhakti AND Madabhushi, Anant},
title = { { Single-Model Lesion Detection and Segmentation for Moderate–Severe Traumatic Brain Injury on T1-weighted MRI } },
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}
}
