A Lightweight and Effective nnU-Net Framework for Whole-Body Pan-Cancer Segmentation

Hui Meng, Lusha Li, Pengyu Zhu, Yinlong Wang, Hongyu Liu, Zhe Chen

Keywords: Pan-Cancer Segmentation, nnU-Net, Deep learning

TL;DR: a more streamlined and effective approach for whole-body pan-cancer segmentation based on nnU-Net

Abstract: Despite recent advancements medical image segmentation, achieving robust pan-cancer segmentation across diverse anatomical regions remains a significant challenge. To address challenges such as tumor diversity, data heterogeneity, and high computational demands, we explored a more streamlined and efficient approach for whole-body tumor segmentation based on nnU-Net. Specifically, we designed a lightweight 4-layer network architecture, optimized patch size and spacing during preprocessing, and developed a fine-tuned training strategy with extended epochs. Our method achieved an average score of 53.48% and 43.84% for the lesion DSC and NSD on the MICCAI FLARE 2025 challenge public validation dataset and the average running time and area under GPU memory-time curve are 17.8s and 64115MB, respectively.

Submission Number: 9