Abstract:
Objective To develop a few-shot image retrieval workflow addressing the lack of links between textual records and physical images of hospital surgical instruments, and to evaluate category retrieval with limited galleries.
Methods A self-built dataset of 487 images from five categories and a public dataset of 3840 images from six categories were used. A frozen DFN5B-CLIP visual backbone combined features from original images and anchor-region images and ranked categories using nearest-instance and class-prototype similarities. Fusion weights were selected on the self-built development subset and then locked. Ten nested random gallery samplings were conducted on each dataset to compare retrieval with one, three, or five images per category.
Results With five images per category, Recall@1 was (99.62±0.45)% on the self-built dataset and (91.69±2.88)% on the public dataset; class-level mAP on the public dataset was (95.45±1.51)% (mean±standard deviation). Under the same configuration, DFN5B-CLIP had higher mean values for all three retrieval metrics than the other three backbones.
Conclusion The method provides candidate category rankings with limited galleries and technical support for instrument image queries and manual review.