Context-Aware Cutmix is All You Need for Universal Organ and Cancer Segmentation

Due to its important potential for various clinical applications, universal organ and cancer segmentation has attracted increasing attention recently. However, its performance is largely hindered due to issues such as (1) partial and noisy labels from different sources and (2) tremendously heterogeneous tumor cases. In this paper, we propose a novel partially supervised segmentation framework by introducing the \(merge-max\) operation for hard mining among the unlabeled classes. Besides, to take full advantage of the expertly annotated tumor data, we design a novel context-aware CutMix scheme to dynamically perform tumor augmentation during training. We also introduce a useful data-cleaning strategy for self-training and adjust the nnU-Net framework for better efficiency. The average scores of organ DSC, organ NSD, tumor DSC and tumor NSD on the public validation set are 92.18%, 96.33%, 46.26% and 38.65%, respectively. And we achieve scores of 93.17% (organ DSC), 96.76% (organ NSD), 61.49% (tumor DSC) and 49.9% (tumor NSD) on the official test set. The average inference time is 13.95 s, the average maximum GPU memory is 2823 MB, and the average area under the GPU memory-time curve is 14112. Collectively, we ranked second among all submitted teams. Our code is available at https://github.com/luckieucas/FLARE23.