Enhancing Prostate MRI Segmentation with Co-training and Multi-view Slice Label: An Approach to Reduce Label Dependency

Prostate cancer (PCa) is the second leading cause of death among men, following lung cancer. Additionally, it has the highest incidence rate among men. MR images play a crucial and effective role in detecting lesions within the prostate region. Currently, artificial intelligence methods have demonstrated their potential in enhancing diagnostic accuracy for clinicians. However, challenges such as the high cost of medical image annotation persist. This paper applies a method that entails annotating a limited number of slices from two different views (e.g., transverse, sagittal) within the 3D image. Subsequently, the annotations are propagated to the entire 3D image through image registration, facilitating the completion of the prostate segmentation using multi-view co-training. The co-learning segmentation network utilizes both CNN and Transformer architectures to incorporate local and global features into the model. Our method outperforms several classical semi-supervised learning methods when evaluated on a combined dataset comprising ProstateX and TongjiProstate.