Abstract

Brain metastases exhibit substantial variation in number, size, morphology, and magnetic resonance imaging appearance, making automated multi-region segmentation challenging. Postoperative resection cavities are particularly difficult to delineate because of their heterogeneous appearance and limited representation in the training data. This work presents a region-aware framework for the BraTS 2026 Brain Metastases Segmentation Challenge. The framework is implemented within nnU-Net v2 and uses a Residual Encoder U-Net backbone together with a customized asymmetric Tversky–Binary Cross-Entropy objective for the simultaneous prediction of whole tumor, tumor core, enhancing tumor, and resection cavity regions. During inference, probability maps from five independently trained folds are averaged using probability-space ensembling before applying region-specific probability thresholds, logical region-consistency rules, and connected-component filtering. The framework was evaluated using the official BraTS 2026 validation platform. Relative to the submitted standard nnU-Net baseline, the proposed configuration achieved higher Instance F1, lesion-wise Dice, normalized surface Dice, and lower 95th-percentile Hausdorff distance for all four evaluated regions. The largest improvement was observed for resection cavity segmentation, for which Instance F1 increased from 0.238 to 0.586, lesion-wise Dice increased from 0.459 to 0.578, and HD95 decreased from 137.59 mm to 88.10 mm. Improvements were also obtained for enhancing tumor, tumor core, and whole tumor segmentation. These results indicate that region-aware optimization combined with probability-space ensembling and inference-time refinement can improve multi-region brain metastasis segmentation on the official challenge validation set.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_METS_008.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/forum?id=swAAtfR82q

BibTex

@InProceedings{BolSri_RegionAware_MICCAISAT2026,
        author = { Bollepalli, Sri Sai Hemanth},
        title = { { Region-Aware Residual Encoder nnU-Net for Multi-Region Brain Metastasis Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17253},
        month = {pending},
        page = {pending}
}


back to top