AI Triage for Bone Scans Reaches High Internal Accuracy in Oncology
Whole-body bone scintigraphy remains a workhorse test in oncology because osseous metastases can be widespread, subtle, and clinically consequential. The challenge is not the availability of scans; it is the volume of images and the variability in how disease appears from one patient to the next. A deep learning system named OsseoScan AI was built to classify whole-body bone scans as normal or abnormal, and its performance suggests a realistic role as a screening layer before a physician reads the study in full.
Investigators trained the model on anterior and posterior whole-body images from a single center in Thailand, using scans acquired from April 2021 through June 2026. The dataset included 5,040 scans from 4,313 patients, of which 1,395 were abnormal. The work used a two-tier composite reference standard, and the image backbone was an ImageNet-21k-pretrained EfficientNetV2-B3. Patient-level partitioning separated 4,032 scans into a training and validation pool, while 1,008 scans were held out for testing. That design matters because scan-level splitting can leak information across groups, inflating performance; patient-level separation is more defensible for clinical evaluation.
Performance Held Up in Internal Validation, Especially for Extensive Disease
Across five-fold cross-validation, pooled out-of-fold AUC reached 0.959 with a 95% confidence interval of 0.952 to 0.967. On the independent test set, AUC was 0.963 with a 95% confidence interval of 0.946 to 0.977 at the pre-specified threshold. Sensitivity was 0.907, specificity was 0.886, and negative predictive value reached 0.961. In a setting where the practical aim is often to rule out obvious normality or quickly route clearly abnormal studies to the front of the queue, those figures are hard to ignore.
The strongest performance came when disease burden was more obvious. Sensitivity was 1.000 for extensive metastases and 0.925 for multifocal metastases. Solitary lesions were different. Sensitivity dropped to 0.725 in that subgroup, a result that exposes the central weakness of many image classification systems: they do better when signal is diffuse and struggle when pathology is sparse, focal, and easy to miss against normal skeletal uptake.
Solitary Metastases Exposed the Main Diagnostic Weak Point
A solitary metastatic focus on bone scintigraphy can be biologically and visually elusive. Uptake may be faint, anatomically ambiguous, or partially obscured by adjacent physiologic activity and degenerative change. A model trained to distinguish normal from abnormal whole-body images will naturally favor the features that recur most often and most vividly in the training distribution. In practical terms, a scan with multiple lesions offers stronger pattern recognition than a scan with one subtle hot spot.
The study addressed this limitation with a sensitivity-oriented threshold. At that setting, solitary-lesion sensitivity rose to 0.882, but 40.7% of scans were flagged. The tradeoff is classic in diagnostic medicine: greater sensitivity arrives at the cost of a much larger review burden. For a worklist triage tool, that may still be acceptable if the goal is to pull more suspicious scans forward without replacing specialist interpretation. For autonomous decision-making, the false-positive burden remains too high and the missed-lesion risk is still too important to ignore.
Explainability Tools Offered Visual Cues, Not Clinical Certainty
Grad-CAM was used to provide qualitative explainability. In a narrow sense, that helps show which regions influenced the network’s decision. In a broader clinical sense, it does not solve the core problem of trust. Heat maps can make a prediction more interpretable, but they do not prove that the model recognized the correct lesion biology, ruled out mimics, or understood the difference between metastasis and benign osteoblastic activity. Physicians still need to see whether the model is focusing on ribs, spine, pelvis, or other skeletal regions in a way that aligns with real radiologic judgment.
The model’s value, then, is likely as a front-end sorter rather than a final arbiter. A high negative predictive value could help reduce the cognitive load on nuclear medicine readers when scans appear convincingly normal. A high AUC could support prioritization in a busy oncology imaging service. Yet none of those uses escape the need for human oversight, because the highest-stakes failures in nuclear medicine often occur in exactly the kind of borderline case the model handles least well.
Single-Center Retrospective Data Limit Near-Term Clinical Adoption
The cohort came from Phrapokklao Hospital in Thailand, and the study was retrospective and single-center. Internal validation can establish whether a model works on local data that resemble its training set. It cannot establish whether the system will hold up across different camera systems, reconstruction settings, reader practices, tumor mixes, or patient populations. External, multicenter, and prospective evaluation is still required before clinical use.
Ethics approval came from the Phrapokklao Hospital Ethics Committee, with waiver of individual informed consent because of the retrospective design. The study was conducted in accordance with the Declaration of Helsinki, and the authors reported no competing interests. Funding came from the Medical Education Center at Phrapokklao Hospital.
For oncology services, the most immediate implication is not replacement of interpretation but possible triage support. A model that can reliably separate obviously normal scans from likely abnormal ones may shorten turnaround in some settings, especially where the volume of staging scans is high and staffing is limited. The unresolved question is whether performance remains strong once the system faces new institutions, new patient mixes, and the smaller, harder lesions that drive the most consequential misses. Until that evidence exists, OsseoScan AI should be viewed as a promising internal prototype for worklist prioritization, not a finished clinical tool.