Abstract
We present our submission to the AIMS-TBI 2026 Challenge for automated detection and segmentation of traumatic brain injury (TBI) lesions from T1-weighted MRI. We train two architecturally distinct 3D segmentation networks - a custom lightweight 3D U-Net (~6.5M parameters) and a MedNeXt-S model (~5.6M parameters) - on 472 cases under identical loss and optimization. MedNeXt-S substantially outperforms the U-Net baseline, doubling mean Dice from 0.240 to 0.509 on an internal 79-case validation set. Candidate rankings differed by task: segmentation achieved its best point estimates with a heterogeneous probability ensemble (25% U-Net + 75% MedNeXt, mean DSC 0.558, HD95 22.3 mm on an organizer-provided secondary validation cohort of 102 cases), while detection achieved highest point-estimate balanced accuracy with MedNeXt alone (0.845, 95% CI [0.774, 0.907]). On the secondary cohort, the detection point estimate favored MedNeXt alone, but the paired comparison was inconclusive (McNemar p = 0.45; delta balanced accuracy = 0.029, 95% CI [-0.021, 0.081]). Because this cohort was used for final configuration selection, these estimates carry selection bias and unbiased assessment is deferred to the hidden challenge test set.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AIMS_TBI_005.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=~Shohei_Yoshimoto1
BibTex
@InProceedings{YosSho_TaskSpecific_MICCAISAT2026,
author = { Yoshimoto, Shohei},
title = { { Task-Specific MedNeXt and Heterogeneous Ensembling for Traumatic Brain Injury Lesion Analysis 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}
}