A Zero-Shot Learning Framework for Accurate Appendix Segmentation Based on SAM (AAZ-SAM)

The deep learning-based automatic segmentation of the appendix CT images faces the challenge of scarce annotations, necessitating the development of high-precision segmentation methods with minimal or no reliance on annotated data. To address this, we propose a zero-shot learning framework for appendix segmentation based on the segmentation large model Segment Anything Model (SAM), named AAZ-SAM, which consists of two modules: 1) the Guidance Information Expansion Module (GIEM) generates complete layer prompt information through linear interpolation of three key layer prompt boxes (first, middle, and last layers), eliminating the need for layer-by-layer annotated prompt boxes, thereby improving clinical efficiency. 2) the Feature Focusing Module (FFM) utilizes Gaussian focusing to concentrate the model’s attention, enhancing segmentation performance without the need for model fine-tuning, achieving an improvement in the DSC from 0.817 to 0.836 on the internal test sets, and 0.634 to 0.668 on the external test sets. This study demonstrates that the innovative strategies can improve the performance of large models in medical imaging, enhancing the efficiency and segmentation accuracy of SAM in appendix segmentation tasks.