Abstract
In this paper, we propose Adaptive TransVNet (ATVNet), a transformer-enhanced VNet-based framework for automated moderate-to-severe traumatic brain injury (msTBI) lesion segmentation from T1-weighted MRI images. Unlike better-characterized neurological conditions such as stroke, multiple sclerosis, or brain tumors, msTBI lesions exhibit substantial variability in size, location, and appearance, making automated segmentation a challenging task. The proposed architecture extends the conventional V-Net framework by incorporating transformer modules at multiple low-resolution stages and introducing learnable residual gating to adaptively regulate the contribution of transformer features. The gating parameters are initialized to preserve the original convolutional network behavior and gradually learn the optimal integration of local and global representations during training. The model was developed and evaluated within the AIMS-TBI26 challenge, using extensive three-dimensional data augmentation and a combined BCE-Dice-Focal loss function. Preliminary experiments demonstrated improved performance over both V-Net and TransVNet variants, with ATVNet achieving a mean DSC of 56.52% on the benchmark train-validation split. A 5-fold ensemble strategy was also evaluated but did not improve performance for either task. In the final challenge evaluation on a hidden test dataset, the submitted model achieved a DSC of 52.25%, an ASSD of 9.12 mm, and an HD95 of 32.02 mm for lesion segmentation. For lesion detection, the proposed approach achieved a balanced accuracy of 82.89%, with 75.00% sensitivity and 90.79% specificity. These results demonstrate the potential of the proposed approach within the AIMS-TBI26 challenge, while further evaluation on larger, more diverse datasets and additional performance metrics is needed to fully assess its general applicability.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AIMS_TBI_003.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=~Bal%C3%A1zs_Zavadil1
BibTex
@InProceedings{ZavBal_ATVNet_MICCAISAT2026,
author = { Zavadil, Balázs AND Szabó, Bálint AND Szlávecz, Ákos AND Benyó, Balázs},
title = { { ATVNet: Adaptive Transformer-Enhanced V-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}
}