Efficient Pan-Cancer Lesion Segmentation from Partially Labeled Data with nnU-Net

Yannick Kirchhoff, Maximilian Rouven Rokuss, Benjamin Hamm, Ashis Ravindran, Constantin Ulrich, Klaus Maier-Hein, Fabian Isensee

Keywords: FLARE challenge, Pan-Cancer Segmentation, nnU-Net

Abstract: Accurate segmentation of cancer lesions in whole-body CT scans is essential for diagnosis and treatment planning. However, this task is challenging due to the diversity of lesion appearances and sizes, as well as the prevalence of partially labeled datasets. To address these challenges, Task 1 of the FLARE 2024 challenge was launched to encourage researchers to develop algorithms capable of generalized pan-cancer segmentation from a large, partially labeled dataset. In this paper, we describe our contribution to this challenge, utilizing nnU-Net with large batch size training and inference optimizations for efficient segmentation. Our best method achieved an average Dice Similarity Coefficient (DSC) of 15.6\% and an average Normalized Surface Dice (NSD) of 17.3\% on the validation set, with a mean inference time of 71.8 seconds and an area under the VRAM-time curve of 427,572 MB. Our second-best method achieved an average DSC of 13.5\% and an average NSD of 13.8\%, with a mean inference time of 44.9 seconds and an area under the VRAM-time curve of 224,872 MB. These results highlight the significant challenges inherent in pan-cancer lesion segmentation from partially labeled data under resource constraints, and underscore the need for further research in this area.

Submission Number: 20