Semi-supervised Two-Stage Abdominal Organ and Tumor Segmentation Model with Pseudo-labeling

In real-world scenarios, such as abdominal organ and tumor segmentation, obtaining complete labels for all classes often presents significant challenges. Additionally, optimizing GPU efficiency emerges as another critical factor in the abdominal organ and tumor segmentation process. To address the challenge of partial labeling, a semi-supervised approach was employed. Initially, a larger model was trained using complete labels, which was then utilized to generate pseudo-labels. Subsequently, a smaller model was trained on these pseudo-labels. To mitigate GPU memory consumption, a two-stage strategy was implemented. Firstly, an abdomen location model was trained to accurately identify the abdominal area. Subsequently, the segmentation process was restricted to this localized area, thereby reducing the GPU memory requirements. Experiments on the FLARE23 challenge exhibited promising performance, with an average actual running time of 25.971 s, an average AUC-GPU (Area Under the Curve of GPU memory consumption) of 28463.7 MB, and an average maximum GPU memory usage of 2.6 GB on the validation set, and the average running time on the testing set was 18.95 s, with AUC-GPU of 20790 MB. Moreover, the model achieved a Dice coefficient (DSC) of 79.99% for organ segmentation and 27.99% for tumor segmentation on the public validation dataset, and 80.84% and 24.26% for the DSC of organ and tumor segmentation on the testing result.