ProtoMGNet: Memory Guided Network With Instance Prototype Enhancement for Echocardiography Segmentation

Accurate segmentation of key cardiac structures in echocardiography is crucial for early diagnosis of cardiovascular diseases, yet remains challenging due to inherent imaging limitations including low signal-to-noise ratios, speckle noise, and viewpoint variations. Existing methods focus on mining intrinsic image information to optimize feature representation, and their performance deteriorates greatly under these challenging imaging conditions. In contrast, cardiologists are capable of achieving accurate identification by leveraging accumulated clinical experience. We posit that the core of this ability stems from the integration of prior knowledge stored in memory with observed ambiguous structural patterns and local visual cues-such as textures and edges-to reconstruct a comprehensive representation of the cardiac anatomy. Inspired by this cognitive memory reconstruction mechanism, the ProtoMGNet, consisting of Prototype Enhanced Memory Reconstructor (PEMR), Texture Feature Mixer (TFM), and Frequency Domain Edge Filter (FEF), is proposed with the aim of leveraging experiential memory to improve segmentation accuracy. Specifically, the PEMR introduces a dynamically updated memory unit storing the dataset-level distribution information of the class and utilizing predicted masks for weighted aggregation to reconstruct memory representation. Additionally, it incorporates instance prototypes for memory matching to enhance the discriminative nature of reconstructed features. The TFM and the FEF mix shallow texture information and high-frequency weak edge features, respectively, to further optimize the segmentation boundaries. Experiments on public datasets CAMUS and CardiacUDA show that ProtoMGNet outperforms 13 state-of-the-art methods, showing great potential as a clinical auxiliary tool.