SGDA-Net: Interpretable Medical Image Segmentation via Second-Order Weighted Graph Convolution and Dynamic Agent Attention

Abstract

Medical images constitute a class of highly complex visual multimedia data, where accurate segmentation heavily relies on the synergistic modeling of global semantics and fine-grained local details. However, existing deep learning methods often fall short in depicting complex geometric morphologies and facilitating efficient feature interactions, while suffering from the inherent explainable bottleneck of ''black-box'' decision making. To address these issues, we propose SGDA-Net, a novel segmentation network that integrates second-order weighted graph convolution and dynamic agent attention within an asymmetric encoder-decoder architecture. In the decoder, we introduce a second-order weighted graph convolution that explicitly models second-order curvature, enabling precise reconstruction of boundary structures and complex geometric morphologies. Complementing this, a dynamic agent attention module is designed to employ a saliency-guided strategy to adaptively focus on key regions with high feature activation levels. This synergistic mechanism, based on geometric curvature perception and saliency-driven feature selection, not only achieves efficient fusion of global and local features but also provides feature-level transparency for network decisions. Consequently, our method effectively enhances computational explainability by offering intuitive visual cues for complex clinical reasoning. Extensive experiments on two public benchmarks demonstrate that SGDA-Net yields average Dice coefficients of 85.44% on the Synapse abdominal multi-organ CT dataset and 92.44% on the ACDC cardiac MRI dataset, outperforming the existing state-of-the-art methods with particularly significant performance gains on challenging, small-volume, and morphologically complex targets (e.g., gallbladder and pancreas). Our work illustrates how combining geometry-aware and attention-driven mechanisms can advance both the performance and transparency of understanding medical multimedia content.