Open-Set Recognition for Dental Implant System Identification: A Multicenter Retrospective Study

Abstract Background Deep learning models for dental implant identification are typically trained on a fixed set of systems, although new systems continuously enter the market. Such closed-set models assign unregistered implants to known systems, often with high confidence. We compared open-set recognition methods for detecting unregistered implant systems while identifying registered ones. Methods In this retrospective study, 100,105 implant images from panoramic radiographs across 15 institutions were obtained from a public dataset. Thirteen systems with > 2,000 images were defined as known classes, while 29 systems (14,004 images) were withheld as unknown classes. A ConvNeXt classifier was trained using ten implant-level splits. Score-based, distance-based (OpenMax), and training-based (ARPL) methods were compared at a threshold accepting 95% of known validation images. Results The classifier achieved 0.901 accuracy for system identification and 0.968 accuracy at the manufacturer level. OpenMax achieved an AUROC of 0.898 and detected 61.4% of unknown images with 0.868 precision, outperforming all other methods across ten splits (AUROC, 0.825–0.844; detection, 29.1–49.4%; adjusted P = 0.012). Its advantage was retained after excluding institutions without known-class images. Unknown systems from manufacturers represented in training were harder to detect than those from new manufacturers (57.5% vs. 64.4%), and undetected implants were often assigned to a small number of known systems, frequently from the same manufacturer. Conclusions Closed-set evaluation overlooks a major risk of automated implant identification. A post hoc distance-based method detected most unregistered implants without retraining and with minimal loss of closed-set accuracy, providing a practical safeguard for existing classifiers and supporting expert review of flagged cases.