Ontology-regularized hierarchical transformers for anatomy and phenotype-aware medical image retrieval

Medical image retrieval (MIR) is an essential tool for diagnosis, clinical research, and education, facilitating radiologists’ access to semantically similar, prior cases. However, most deep learning–based MIR systems predominantly focus on visual similarity, neglecting the extensive domain knowledge encapsulated in standardized clinical ontologies. This oversight frequently results in retrievals that are visually similar yet semantically inconsistent, thereby diminishing their diagnostic utility and generalization capability across datasets. To address these challenges, we introduce an ontology-regularized hierarchical transformer framework that explicitly integrates anatomical and phenotypic knowledge into the MIR pipeline. Our model utilizes a multi-scale hierarchical vision transformer to extract fine-to-coarse image features, which are subsequently aligned with an ontology graph encoder derived from RadLex, the Foundational Model of Anatomy (FMA), and the Human Phenotype Ontology (HPO). We propose a graph-aware contrastive loss combined with a hierarchical margin term to draw embeddings of ontologically related concepts closer together, while distancing unrelated ones. Evaluation was conducted on two large-scale public datasets, MIMIC-CXR as the source domain and VinDr-CXR for cross-dataset testing, with additional experiments on DeepLesion for modality variation. The results indicate that our approach achieves significant improvements in retrieval quality, with a +17.3% enhancement in expert-judged semantic correctness and higher cross-dataset recall compared to robust transformer baselines. We further propose an Ontology-Consistency@k metric to quantify the semantic alignment of retrievals and demonstrate its correlation with expert ratings. Qualitative analyses revealed enhanced interpretability through concept-attribution maps that localize ontology-relevant areas. This study represents a step toward clinically meaningful, explainable, and generalizable MIR systems built entirely on publicly available data.