Abstract
Brain metastases are typically small, often multifocal, and evaluated by lesion-wise metrics that penalize every spurious predicted component, so that the balance between sensitivity and precision determines much of the achievable score. We present our entry to the BraTS-METS 2026 challenge (Task 1), a two-stage pipeline built on nnU-Net. The first stage is an ensemble of five cross-validation fold models from a single nnU-Net configuration, trained with an extended 5000-epoch schedule and a Dice-plus-TopK cross-entropy loss. The second stage is a learned connected-component false-positive rejector applied exclusively to the enhancing-tumor prediction: each predicted component is described by confidence, shape, and intensity features and scored by a gradient-boosted classifier trained solely on out-of-fold predictions, with a decision threshold derived from a lesion-wise Dice cost-benefit analysis and fixed before submission. In the course of development we evaluated four separate interventions intended to recover additional small enhancing lesions; each detected more true lesions, each also introduced a comparable or larger number of false ones, and each reduced lesion-wise enhancing-tumor Dice. Reversing this direction and removing components instead produced our largest single improvement, raising lesion-wise enhancing-tumor Dice from 0.6964 to 0.7187 and normalized surface distance from 0.7608 to 0.7839 on the challenge validation set, with corresponding gains in tumor core and whole tumor through subregion nesting. The same operation reduces the small-instance F1 score, a trade-off we report and analyze explicitly because both metric families contribute to the challenge ranking. For our system, precision rather than sensitivity was the binding constraint on lesion-wise performance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_METS_029.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{GirHar_Precision_MICCAISAT2026,
author = { Girish, Hari AND Kong, Michael AND Chan, Sherman AND Patil, Nihal AND Chan, Jany AND Chakravarti, Arnab AND Zhu, Simeng},
title = { { Precision Over Recall: Brain Metastasis Segmentation with a Learned False-Positive Rejector } },
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}
}