Universal Medical Image Segmentation with Task-Specific Prompt-Guided Transformer Model
Weilin Luo, Hao Niu, Junjie Hu, Ying Cai, Daji Ergu, Haitao Lan · 2023
Medical image segmentation plays a crucial role in healthcare by requiring precise and efficient models to delineate lesion areas and critical organ regions, reducing harm to vital organ areas during radiation therapy while determining diagnoses, treatment plans, and appropriate medication dosages. In this context, segmentation models using the Vision Transformer (ViT) architecture have demonstrated impressive and powerful performance. However, conventional ViT models often possess limitations in terms of generalizability, as they are often tailored to specific medical images of particular organs. Consequently, adapting them for segmenting diverse organ regions demands model updates or retraining efforts. This paper introduces an innovative approach: a universal ViT segmentation model driven by task-specific prompts. This method combines task-specific features extracted from encoder of a vision Transformer-based segmentation model with trainable universal prompts, enabling its application to a broader range of medical organ segmentation tasks. This forward-looking framework enhances both the flexibility and efficiency of medical image segmentation, catering to diverse organ segmentation requirements within a unified model architecture. The effectiveness and adaptability of our proposed approach are empirically validated across various medical image datasets.