Abstract
Brain metastasis segmentation is scored lesion-wise, which turns it into a detection problem: each lesion counts equally regardless of volume, and every spurious component is charged twice, once as a zero in the Dice average and once in the F1 denominator. Most challenge effort goes into architectures and losses. We find our decisive gains lay downstream, in how probabilities become a label map. An unweighted average of three nnU-Net ResEnc-L variants decoded by hysteresis thresholding raises lesion-wise Dice from 0.560/0.622/0.593 to 0.725/0.766/0.746, with those two inference-time steps contributing over four fifths of the gain and without retraining the constituent models; hysteresis works because it splits the decision of whether a lesion exists from how far it extends, and never inspects component size. We then characterise the residual error: the median peak probability inside our false positives is 0.96 against 0.55 inside the lesions we miss, and seven suppression strategies spanning the decode, the loss and the seed threshold all fail on this system, which suggests that ranking components by confidence is unpromising here.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_METS_035.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=UBtGFsTmgb
BibTex
@InProceedings{KunSou_InferenceTime_MICCAISAT2026,
author = { Kundu, Soumya Snigdha AND Barfoot, Theodore AND Ivory, Marina AND Muhammad, Kabir Hamzah AND Macias, Lorena Garcia-Foncillas AND Shapey, Jonathan AND Vercauteren, Tom},
title = { { Inference-Time Decoding for 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}
}
