Abstract
Automated segmentation of moderate-to-severe traumatic brain injury (msTBI) lesions on T1-weighted MRI remains challenging because multiple lesions with substantial inter-instance size imbalance may coexist within a scan. Conventional nnU-Net foreground oversampling draws anchors from pooled foreground voxels, making lesion sampling probability approximately proportional to lesion volume and consequently underexposing small lesions. We adapt instance-aware foreground sampling to nnU-Net v2 by decomposing the ground-truth mask into connected lesion instances, uniformly selecting an instance, and then sampling an anchor voxel from that instance. This rebalances lesion exposure during training without modifying the network architecture. We evaluated the approach on the AIMS-TBI 2026 dataset using a 3D full-resolution Residual-Encoder U-Net. On the official AIMS-TBI validation set, averaged across five fold-trained models, instance-aware sampling improved mean Dice from 0.592 to 0.601, while reducing HD95 from 22.864 to 21.731 and ASSD from 7.356 to 5.888 compared with conventional foreground sampling. Lesion-wise analysis further showed improved performance for small lesions, with recall increasing from 0.490 to 0.517, F1 from 0.530 to 0.555, and Matched Dice from 0.654 to 0.667 for lesions with volumes <= 1 mL. These results indicate that balancing foreground sampling at the lesion-instance level can improve msTBI lesion segmentation, particularly for small lesions that are underrepresented by voxel-based sampling.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AIMS_TBI_004.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=~Jin_Tae_Kwak2
BibTex
@InProceedings{NohSeu_Rebalancing_MICCAISAT2026,
author = { Noh, Seungha AND Yu, Cheolhee AND Yang, Hyun AND Kwak, Jin Tae},
title = { { Rebalancing Lesion Exposure: Instance-Aware Sampling for 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}
}