Development and validation of an artificial intelligence model for automated classification of whole-body bone scintigraphy in oncology patients

Whole-body bone scintigraphy with Tc-99 m methylene diphosphonate (MDP) is widely used to detect osseous metastases, and artificial intelligence (AI) may support its interpretation. We developed and internally validated OsseoScan AI, a deep learning model classifying whole-body bone scans as normal or abnormal. In this retrospective single-center study at Phrapokklao Hospital, Thailand, consecutive oncologic scans from April 2021 to June 2026 were labeled using a two-tier composite reference standard. An ImageNet-21k-pretrained EfficientNetV2-B3 was trained on full-body anterior and posterior images from 5,040 scans of 4,313 patients (1,395 abnormal), partitioned at patient level into a training/validation pool (4,032 scans; 5-fold cross-validation) and a held-out test set (1,008 scans). Pooled out-of-fold AUC was 0.959 (95% CI 0.952–0.967). On the test set at the pre-specified threshold, AUC was 0.963 (0.946–0.977), sensitivity was 0.907 (0.868–0.944), specificity was 0.886 (0.862–0.910), and NPV was 0.961. Sensitivity was 1.000 for extensive, 0.925 for multifocal, and 0.725 for solitary metastases; a sensitivity-oriented threshold raised solitary-lesion sensitivity to 0.882 while flagging 40.7% of scans. Grad-CAM provided qualitative explainability. OsseoScan AI showed high discrimination in internal validation and may support worklist prioritization; external, multi-center, and prospective evaluation is required before clinical use.