All-in-One Multi-Organ Segmentation in 3D CT Images via Self-Supervised and Cross-Dataset Learning

Accurate segmentation of organs related to breast cancer metastasis in 3D CT images is crucial for clinical applications such as surgical planning, radiation therapy, and personalized treatment strategies. However, the scarcity of annotated datasets poses challenges in training robust models. This work introduces a novel framework combining self-supervised learning (SSL) and cross-dataset label integration to develop an All-In-One (AIO) segmentation model. We pretrain an encoder using contrastive learning on over 6,000 unlabeled CT images, enhancing feature extraction for the segmentation of 6 key organs without annotations. Organ-specific models are trained on individual datasets, and cross-dataset inference generates pseudo labels for unannotated organs. These pseudo labels, combined with ground truth, create a comprehensive training set for the AIO model. Our approach improves the Dice coefficient for segmentation from an average of 89.48% to 91.40%, effectively addressing the challenge of limited annotations. This advancement has the potential to enhance diagnostic accuracy and reduce the workload of imaging specialists.