Partial-Labeled Abdominal Organ and Cancer Segmentation via Cascaded Dual-Decoding U-Net

Zhiyu Ye, Hairong Zheng, Tong Zhang

Abstract: In the FLARE2023 challenge, we developed a cascaded dual-decoding U-Net framework to address the complex task of partial-labeled abdominal organ and cancer segmentation. Initially, we explored the potential of 3D transformer-based models but transitioned to 2D U-Net solutions due to computational resource and inference time constraints. We first trained separate 3D models for cancer and full-organ segmentation using data that included labels for both cancer and full organs. Subsequently, we generated pseudo labels for unlabeled and partially labeled data based on these initial models. To enable a single model to effectively learn and infer both organ and cancer labels within images, we designed a dual-decoding structure based on the 2D U-Net architecture. Our training process involved several steps with various subsets of the training data. By comparing our model trained without unlabeled data, we discussed the impact of unlabeled data and its pseudo labels on the experimental results. Our method, the version trained without unlabeled data, achieved an average DSC score of 83.22% for organs and 33.22% for lesions on the validation set. The average running time and area under the GPU memory-time curve were 33.8 s and 50066.25 MB, respectively. The codes has been open-sourced to https://openi.pcl.ac.cn/OpenMedIA/pclmedia_FLARE23.