Abstract
Brain metastasis segmentation in multi-modal MRI is challenged by two pervasive issues in clinical practice: missing MRI sequences and extreme lesion size heterogeneity. We present a two-stage framework for brain-metastasis segmentation under incomplete acquisitions. In the first stage, missing T2-weighted (T2W) sequences are imputed via a latent-space ensemble of a deterministic modality-translation encoder-decoder (MT-ED) and a Brownian-bridge diffusion model (MT-BBDM), built on a volumetric compression network with tumor-aware VAE fine-tuning; the recovered cases are added to the training set in place of discarding them. In the second stage, a custom nnUNetv2 6-stage 3D U-Net segments brain metastases, combining a $3 \times$ cross-entropy weight on the underrepresented resection-cavity (RC) class with Focal Loss ($\gamma = 2$) to redistribute gradient toward small lesions and inter-class boundaries. On a dataset of 1,296 patients with 9,118 histopathologically confirmed metastatic lesions annotated under the BraTS 2026 Task1 protocol, the model achieves instance-level F1 scores of 0.821 (tumor core), 0.810 (enhancing tumor), 0.801 (whole tumor), and 0.493 (resection cavity) on a 104-case internal test set. Size-stratified analysis reveals a pronounced gap: small lesions (<50 voxels) attain F1 roughly half those of large lesions, and small RC lesions are entirely undetected, underscoring the need for lesion-size-aware modeling. Our code is available at \url{https://github.com/hhhx-lab/ECNU_NYU_brats.git}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_METS_028.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=tUdjB4sdVS
BibTex
@InProceedings{HeJia_Missing_MICCAISAT2026,
author = { He, Jianbao AND Sun, Zhuoran AND Fan, Boyu AND Zhang, Xueyang AND Zhao, Mingrui AND Shu, Hai AND Chen, Ziqi},
title = { { Missing Modality Imputation and Brain Metastasis Segmentation via Latent-Space Ensemble Synthesis and Size-Aware Focal Loss } },
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}
}