Hypergraph Neural Network-Based Anatomical Relationship Modeling for Complex Disease Pattern Discovery from Medical Images

Abstract

Medical image interpretation for complex diseases remains challenging because conventional deep learning approaches primarily capture local visual features or pairwise anatomical interactions, limiting their ability to represent higher-order structural dependencies among multiple organs and tissue regions. This paper presents the HGNN-ARM (Hypergraph Neural Network-Based Anatomical Relationship Modeling) framework that considers anatomical regions as hypergraph nodes and anatomical relations as hyperedges for the purpose of discovering patterns of complicated illness in medical images. Furthermore, integration of multi-scale feature extraction, hypergraph convolution-based learning, attention-based aggregation of anatomical relations, and disease classification is an important part of this framework that allows creation of specific features for analysis of various anatomical regions. Evaluation of the model under study featured one of the well-known datasets with medical images, consisting of over 8,400 images for six complex disorders divided into the following subsets: 80% for training, 10% for validation, and 10% for testing. The performance of the model is assessed based on various criteria, including accuracy, precision, sensitivity, F1-score, area under the ROC (Receiver Operating Characteristic) curve, and Matthew’s correlation coefficient. In the course of the evaluation, the proposed HGNN-ARM model reached 98.12% accuracy, 97.84% precision, 97.65% recall, 97.74% F1-score, 0.992 AUC, and 0.964 MCC. According to the statistical analysis performed, the improvements in performance seen were statistically significant (p < 0.001). In addition, results from the ablation study indicate that multi-scale feature extraction, hypergraph-based modeling of anatomical relationships, and attention-based aggregation of the relationships were important factors. Based on these findings, it can be concluded that higher-order anatomical relationship learning offers an effective framework for computer-aided diagnosis of complex diseases.