SemiOrth: a novel orthogonal dual network architecture for enhanced semi-supervised medical image segmentation

Medical image segmentation is crucial for Computer-Aided Diagnosis (CAD) in healthcare. However, the confidentiality of medical data and the difficulty of creating segmentation annotations require training in high-precision models with limited annotated data. This study introduces a novel co-model training framework that uses an orthogonality loss function to enhance the heterogeneity between two models, enabling them to focus on different target features more effectively. Additionally, we employ a hybrid approach to refine pseudo-labels by evaluating predictions from the same image under varying augmentations, improving their accuracy. Tested on the Automated Cardiac Diagnosis Challenge (ACDC) dataset, our methodology achieved competitive performance with only 8% labeled data, underscoring its efficacy in optimizing medical image segmentation with minimal annotations.