Abstract
Brain metastases are the most frequently diagnosed intracranial tumors in adults, and their accurate detection and volumetric segmentation are essential for treatment planning and response assessment. This task is challenging because metastatic lesions are often small and numerous, making them easy to miss while also being prone to false-positive predictions. We address the BraTS 2026 Brain Metastases challenge with a pipeline that couples a segmentation branch and a detection branch. The segmentation branch, built on nnU-Net, comprises a region-based primary model and a dedicated resection-cavity (RC) specialist. Their predictions are combined by a priority-based merging rule to produce the multi-class segmentation, and a size-based post-processing then removes connected components smaller than 5 mm³. In parallel, the detection branch, an nnDetection model trained on the enhancing tumor (ET), provides an independent second opinion. A detection-guided consensus keeps a small enhancing-tumor component (smaller than 27 mm³) only when its centroid falls within a detection box, suppressing spurious small detections while preserving lesions corroborated by the detector. On the challenge validation set, our full pipeline (configuration 5) achieved lesion-wise Dice similarity coefficients (DSC) of 0.745, 0.671, 0.765, and 0.720 for the ET, RC, tumor core (TC), and whole tumor (WT), respectively, with corresponding normalized surface distances (NSD) of 0.808, 0.547, 0.816, and 0.735. Relative to the segmentation-only configurations, using the detector to prune false-positive small detections yielded the largest improvement in these overlap metrics, at small cost to the small-lesion detection F1-score (e.g., 0.501 for ET). Our code is available at https://github.com/deepnoid-ai/DEEP-MET.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_METS_005.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{KimJun_Integrating_MICCAISAT2026,
author = { Kim, Jun-Ho AND Nam, Ki Bok},
title = { { Integrating nnDetection and nnU-Net for Brain Metastasis Segmentation in the BraTS-METS 2026 Challenge } },
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}
}
