Abstract
The generalizability of brain tumor segmentation models is an important clinical factor in the deployment of deep learning based models in hospital settings. Usually, the segmentation models are trained and evaluated on a single tumor type, yet clinical adoption requires robustness across different tumor data cohorts that differ in age, geography, imaging protocols, and scanner, etc. The BraTS Generalizability Across Tumors (GoAT) task evaluates this setting on images from multiparametric magnetic resonance imaging, where training data has glioma, meningioma and metastasis tumor types while the held-out validation dataset includes pediatric and sub-Saharan African cases. We proposed RACER (Radiomic-Adaptive Cohort Ensemble with Refinement), a deep learning based methodology, for this task. In the proposed methodology, three convolution neural network (CNN) based encoder were trained as backbones, two of which with residual encoder were conditioned per case through a radiomic feature-wise linear modulation layer, and a self-supervised masked-image-modelling initialization together with cohort-shift augmentation was used to encourage generalization. The five-fold softmax ensemble was refined by a dual-encoder network that corrects the ensemble prediction through a zero-initialized residual, and a cohort-safe relabeling rule was applied as post-processing. On the official validation synapse portal, the final model reached a mean Dice of 0.840, a normalized surface Dice of 0.497 and a mean 95th percentile Hausdorff distance of 21.3 mm, and a sub-region mean Dice of 0.812, 0.830, 0.877 for enhancing tumor, tumor core, whole tumor, respectively, thus improving over the ensemble on all three region-averaged metrics. (Source code: https://github.com/piyushkp27/RACER)
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_GoAT_008.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{PraPiy_RACER_MICCAISAT2026,
author = { Prajapati, Piyush Kumar AND Sheikh, Mohammad Tufail AND Choudhary, Arpit AND Parikh, Megh AND Singh, Anup},
title = { { RACER: A Radiomic-Adaptive Cohort Ensemble with Refinement for Cross-cohort Brain Tumor Segmentation using Deep Learning } },
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}
}