Domain-Adaptive Self-Supervised Contrastive Hashing for Scalable Chest Radiograph Retrieval
IEEE Access · 2026 · 14 · 113690-113707
DOI: 10.1109/access.2026.3713435 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Shulong Pan. Die Arbeit nennt 63 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
The rapid growth of digital chest radiograph archives creates a need for retrieval systems that are accurate, compact, and adaptable across heterogeneous clinical datasets. Current methodologies often compromise retrieval accuracy for computational efficiency or rely heavily on extensive manual annotation to accommodate domain-specific nuances. We propose a domain-adaptive self-supervised contrastive hashing framework for chest radiograph retrieval that integrates four components: 1) self-supervised contrastive pre-training of a Vision Transformer, utilizing anatomically informed augmentations to capture radiographic-specific invariances; 2) parameter-efficient domain adaptation through low-rank adapter modules, which align the pre-trained backbone with diverse chest X-ray datasets without full model fine-tuning; 3) weak supervision via pseudo-labels automatically extracted from free-text radiology reports; and 4) an end-to-end learnable hashing head optimized jointly for quantization fidelity and retrieval-aware triplet objectives, producing compact 64-bit binary codes. We evaluate on NIH ChestX-ray14, MIMIC-CXR, and PadChest using patient-level query gallery splits and an explicit label-overlap relevance rule, reporting Precision@1, 5, 10, mean Average Precision, and Recall@5. On the three datasets the proposed method attains significant values on Precision@1 and mAP, consistently improving over ResNet-50 and PCA-reduced embedding baselines, while retaining sub-20 ms query latency and a 40 MB index footprint for 200k images. The evaluation additionally specifies protocols for component-wise ablation, pseudo-label quality assessment, robustness under clinically plausible perturbations, domain-mismatch analysis, and computational-cost reporting. These results demonstrate the evidence for retrospective chest radiograph retrieval rather than general medical-image retrieval or clinical deployment readiness.
Zitieren
APA
Pan, S. (2026). Domain-Adaptive Self-Supervised Contrastive Hashing for Scalable Chest Radiograph Retrieval. IEEE Access, 14, 113690-113707. https://doi.org/10.1109/access.2026.3713435
Vancouver
Pan S. Domain-Adaptive Self-Supervised Contrastive Hashing for Scalable Chest Radiograph Retrieval. IEEE Access. 2026;14:113690-113707. doi:10.1109/access.2026.3713435
BibTeX
@article{Pan2026,
author = {Pan, Shulong},
title = {Domain-Adaptive Self-Supervised Contrastive Hashing for Scalable Chest Radiograph Retrieval},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {113690--113707},
doi = {10.1109/access.2026.3713435}
}
EUVIMED