Abdominal Organs and Pan-Cancer Segmentation Based on Self-supervised Pre-training and Self-training

Despite the effective progress in automatic abdominal multi-organ segmentation methods based on deep learning, there are still few studies on general models for abdominal organ and pan-cancer segmentation. Additionally, the manual annotation of organs and tumors from CT scans is a time-consuming and labor-intensive process. To deal with these problems, an efficient two-stage framework combining self-supervised pre-training and self-training is proposed. Specifically, in the first stage, we adopt the Model Genesis method for image reconstruction to promote the model to learn effective anatomical representation information, thereby improving the model's perception of anatomical structures in downstream segmentation tasks and generating high-quality tumor pseudo-labels. Afterward, we fuse partial organ fine-standard of labeled data with pseudo-labels to improve the organ labeling quality. In the second stage, we overlay the generated tumor pseudo-labels onto the corresponding regions of the organ pseudo-labels, and the final pseudo-label images are used to train the nnU-Net model for efficient inference. The proposed method has been evaluated on the FLARE2023 validation cases, and get a relatively good segmentation performance. The average DSC and NSD for organs are 91.51% and 95.52%, respectively. For tumors, the average DSC is 43.47%, and the average NSD is 33.81%. In addition, the average running time and area under the GPU memorytime curve are 85.4 s and 246157.2 MB, respectively. On the test set, we achieved average organ and tumor DSC of 92.17% and 54.99%, respectively, and average inference time of 95.83 s. Our code is publicly available at https://github.com/lihe-CV/HiLab_FLARE23