DAGM-Net: A Dynamic Adaptive Graph and Multi-scale Network for Accurate Jaw Cyst Segmentation

Accurate diagnosis and segmentation of jaw cysts hold substantial clinical significance; however, current deep learning approaches often struggle to capture complex anatomical structures and handle contextual variability. To address these challenges, this paper introduces a novel Dynamic Adaptive Graph and Multi-scale Network (DAGM-Net), which incorporates several innovative layers designed to enhance segmentation performance. Firstly, the Dynamic Graph Topology Learning (DGTL) layer adaptively constructs graph connectivity based on node feature similarity, enabling the model to better capture semantic relationships within the data. Next, the Residual Graph Convolution with Feature Rectification (RGC-FR) layer propagates node information through a three-stage feature rectification process, effectively compensating for feature loss and improving discriminative representation. Additionally, the Progressive Multi-scale Aggregation (PMA) layer hierarchically fuses multi-scale encoder features, thereby enriching contextual information and increasing representational power. To further strengthen optimization, a unified loss with momentum-based adaptive weighting is employed to dynamically balance multiple objectives and promote stable training. Comprehensive evaluations on benchmark datasets demonstrate that DAGM-Net achieves effective improvements in jaw cyst segmentation compared to existing methods.