Dual-Guided Prototype Alignment Network for Few-Shot Medical Image Segmentation

Given the hurdles of limited data availability, annotation challenges, and constrained generalization capabilities in medical image segmentation, few-shot learning (FSL) has become a prominent approach due to its efficacy in learning new categories from minimal annotated samples. Existing methods predominantly adopt the support-guided paradigm, wherein support prototypes, derived from annotated support images, guide the segmentation of unlabeled query images. This approach, however, is prone to local information loss and classification bias. To counter these challenges, we introduce a novel dual-guided prototype alignment network (DGPANet) that integrates a bidirectional segmentation architecture. The support-to-query branch incorporates query guidance information into the conventional support-guided approach for segmenting query images, whereas the complementary query-to-support branch segments support images in reverse with the dual guidance provided. This design not only optimizes our network but also alleviates segmentation biases pertaining to support category objects. Furthermore, we propose an adaptive prototype module (APM) to foster information amalgamation by adaptively clustering analogous elements. Moreover, we propose an adaptive prototype alignment loss to better improve the dual optimization (DO) procedure. Extensive experiments show that our proposed DGPANet favors against state-of-the-art (SOTA) approaches on three public datasets, including abdominal-computed tomography (CT), abdominal-magnetic resonance imaging (MRI), and Cardiac-MRI. Codes are at
https://github.com/shiyi0306/DGPANet
.