Abstract
We present task-specific deep learning pipelines for the AIMS-TBI 2026 Challenge, addressing binary lesion detection and voxel-level segmentation from T1-weighted MRI with separate models optimized for each task objective. For segmentation, a MultiTalentV2-pretrained teacher is fine-tuned on original T1 volumes without skull stripping and distilled into a lightweight student network. For detection, we apply an ICV masking cascade in which the detection network operates on ICV-masked T1 volumes, improving specificity on lesion-free scans. The lesion segmentation and detection models use combined Dice and cross-entropy losses, with the segmentation model additionally incorporating a soft-target distillation objective. On the 103-case official AIMS-TBI validation set (55 lesion-positive, 48 lesion-free), the segmentation model achieves overall Dice 0.692 [95% CI: 0.621, 0.759] and lesion Dice 0.551 [0.468, 0.627] (HD95 25.7 mm, ASSD 6.6 mm); the detection cascade achieves balanced accuracy 0.879 [0.818, 0.935] (sensitivity 0.800, specificity 0.958). On the final held-out test set, the segmentation model achieves lesion Dice 0.5243 (HD95 29.5 mm, ASSD 9.3 mm) and the detection cascade achieves balanced accuracy 0.8482 (sensitivity 0.894, specificity 0.803).
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AIMS_TBI_012.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=~Dahyun_Kim4
BibTex
@InProceedings{KimDah_TaskSpecific_MICCAISAT2026,
author = { Kim, Dahyun AND Won, Jongjun},
title = { { Task-Specific T1 MRI Pipelines for TBI: Knowledge-Distilled Segmentation and ICV-Cascaded Detection } },
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}
}
