A bone-scan AI found most abnormal studies—but solitary metastases were its weak point
A Thai hospital model reached 0.963 AUC on a 1,008-scan internal test set. Sensitivity fell to 72.5% for solitary metastases, and no external or prospective evaluation has been completed.
The short version
- What happened
- Researchers retrospectively assembled 5,040 whole-body bone scans from 4,313 oncology patients at one Thai hospital; 1,008 scans were held out at patient level for testing.
- Who it affects
- A prioritisation tool could help an urgent scan reach a specialist sooner, but a solitary lesion is also the pattern most likely to be missed. No patient should be cleared or delayed on this score alone.
- What remains unproven
- Retrospective study
Keep in mind
Retrospective study
Single hospital
Internal validation only
Composite reference standard
72.5% solitary-metastasis sensitivity
No prospective workflow
No patient outcomes
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
Useful for prioritising a worklist—not clearing a patient
A deep-learning system distinguished abnormal from normal whole-body bone scans with an area under the curve of 0.963 on a held-out internal test set, according to a peer-reviewed Scientific Reports paper published on 29 October. At the prespecified threshold, sensitivity was 90.7%, specificity 88.6% and negative predictive value 96.1%. Those figures support investigation as a worklist-prioritisation aid; they do not show that the model can diagnose or exclude bone metastasis without a nuclear-medicine reader.
The main caution appeared in the cases where missing disease may matter most. Sensitivity was 100% for extensive metastases and 92.5% for multifocal disease, but 72.5% for solitary metastases. A sensitivity-oriented threshold raised the solitary-lesion result to 88.2%, yet it would flag 40.7% of all scans for attention. Hospitals would need to decide whether that larger review burden is acceptable before using it to reorder work.[1]
A large internal dataset with a patient-level split
The retrospective study included consecutive oncology scans acquired at Phrapokklao Hospital in Thailand between April 2021 and June 2026. Researchers used 5,040 anterior-and-posterior scan pairs from 4,313 patients, including 1,395 labelled abnormal studies. They kept each patient in only one partition: 4,032 scans entered the training and five-fold cross-validation pool, while 1,008 scans formed a held-out test set. That patient-level separation reduces the risk that repeat scans from one person leak into both training and testing.
Labels came from a two-tier composite reference standard rather than biopsy confirmation for every image. The EfficientNetV2-B3 model was pre-trained on ImageNet-21k, then trained on the full-body scan pairs. Pooled out-of-fold AUC was 0.959 before the held-out result. Grad-CAM heat maps supplied qualitative location cues, but an attention map is not proof that the system used a clinically valid feature or that its highlighted region explains the classification.[1]
The distinction that matters
Why the threshold changes the clinical trade-off
The model can flag more solitary lesions, but doing so sends more scans to the front of the queue.
Prespecified threshold
Classify the held-out 1,008 scans
90.7% overall sensitivity
Solitary disease
Count single-metastasis cases found
72.5% sensitivity
Sensitivity threshold
Lower the alert threshold
88.2% solitary sensitivity; 40.7% flagged
What still has to be demonstrated
Every image came from one hospital, one retrospective workflow and a bounded scanner and patient mix. The model has not been tested at an independent centre, prospectively, or as part of a live reporting service. Its normal-versus-abnormal output also does not identify the cause of an abnormal scan or replace the clinical history, comparison imaging and specialist interpretation that determine what happens next for a patient.
The Medical Education Center at Phrapokklao Hospital supported the work, and the author declared no competing interests. The assessment would strengthen if a preregistered multi-centre study preserved sensitivity for solitary lesions across scanners and cancer types, measured false-alert workload, and showed that prioritisation shortens reporting for urgent scans without delaying normal studies or creating unequal errors. Until then, the evidence supports internal technical performance—not clinical deployment or improved outcomes.[1]