Anatomy-Aware Hierarchical Contrastive Hashing for Efficient Radiograph Classification and Retrieval

Medical-image retrieval systems must balance semantic relevance with storage and search cost while remaining robust to errors introduced by hierarchical routing. This work presents an anatomy-aware hierarchical contrastive hashing framework for radiograph classification and retrieval. A calibrated ConvNeXt-Tiny classifier first estimates anatomical-region probabilities, after which a shared Swin-Tiny encoder and lightweight anatomy-specific heads produce fine-grained predictions and compact binary codes. The hashing objective combines embedding-level supervised contrastive learning with hash-space semantic supervision, route-specific binary prototypes, a sign-margin constraint, quantization, and route-wise bit balance. Confidence-adaptive multi-route database indexing and top-r query routing are used to reduce irrecoverable failures caused by hard Stage-1 assignment. Experimental results on IRMA and MURA datasets reveal that the proposed framework improves retrieval performance and efficiency over competing deep-feature and hashing-based approaches. The Stage-1 classifier on IRMA achieved 97.05% sample-level accuracy, 93.15% macro recall, and 95.33% macro-F1, whereas Stage 2 reported 96% accuracy and 93.2% macro-F1. 128-bit hash codes achieved Precision@20 of 0.96, and mAP of 0.871. On MURA, Stage 1 achieved 96.93% anatomy-classification accuracy, 96.44% macro-F1, and a calibration error of 0.0126. The 14-class Stage 2 model achieved 75.13% accuracy and 74.84% macro-F1. Joint anatomy–abnormality retrieval showed a code-length-dependent trade-off: 32-bit codes obtained the highest mAP of 0.646, while 256-bit codes achieved the best early-rank performance with Precision@10 of 0.631 and nDCG@10 of 0.626. Study-level normal/abnormal prediction achieved an AUROC of 0.860, AUPRC of 0.857, and accuracy of 80.46%. These results support the use of anatomy-aware routing and compact semantic hashing for efficient radiograph retrieval, while also showing that abnormality-level discrimination remains substantially more challenging than anatomical categorization.