METHOD OF DOMAIN KNOWLEDGE INTEGRATION VIA GRAPH NEURAL NETWORKS FOR CARDIAC SEGMENTATION FROM MRI DATA

Abstract

While deep learning has significantly advanced the automated analysis of cardiac magnetic resonance imaging (MRI), a persistent gap remains between high-performance research prototypes and their practical deployment in clinical workflows. In this work, we present a comprehensive, knowledge-integrated method designed to bridge this gap by holistically addressing these operational demands. The proposed method introduces a multi-stage, standards-aware pipeline that transforms raw medical images into auditable, clinically relevant insights. The process begins with a standardized data ingestion module for DICOM and NIfTI formats, ensuring robust de-identification and canonical spatial orientation. Segmentation of cardiac structures is performed by a 3D volumetric model (SKIF-Seg), which is subsequently exported to the ONNX format. This guarantees cross-platform inference portability via ONNX Runtime, enabling consistent performance across diverse computational environments. Crucially, the method moves beyond pixel-level prediction by extracting clinically established measurements, such as ventricular volumes, myocardial mass, and ejection fraction, from the segmentation masks. These metrics are structured as nodes in a knowledge graph, where edges explicitly encode established anatomical and physiological relationships. A Graph Convolutional Network (KI-GCN) then reasons over this structured representation to draw a conclusion, directly integrating domain knowledge into the decision-making process. Evaluated on the public ACDC and M&Ms-2 datasets, our approach achieves competitive segmentation performance with macro-Dice scores of 0,939 ± 0,021 and 0,927 ± 0,025, respectively. The subsequent diagnostic module delivers robust classification results, attaining a macro-ROC-AUC of 0,964 ± 0,018. The key conclusion is that by synergizing domain knowledge with rigorous software engineering practices, our method yields a portable, interpretable, and well-calibrated system that significantly closes the gap between algorithmic performance and true clinical reliability.