The clinical problem at the heart of the study is one of the most persistent frustrations in modern prostate cancer imaging. Prostate-specific membrane antigen positron emission tomography, or PSMA PET/CT, has transformed how doctors stage the disease, offering far better sensitivity and specificity than older techniques such as bone scintigraphy and plain CT. Among the available radioactive tracers, [18F]PSMA-1007 has become especially popular because of its favourable biodistribution: it is cleared through the liver and bowel rather than the urinary tract, allowing a clean view of the prostate and pelvis, and it delivers a high tumour-to-background signal with better spatial resolution and a longer half-life than gallium-68-labelled alternatives.

But [18F]PSMA-1007 carries a well-known trade-off. It produces focal bone uptake that does not correspond to any clear abnormality on the CT component of the scan far more often than other tracers, a phenomenon called unspecific bone uptake. Studies suggest that between roughly 20 and 37 percent of patients have at least one such indeterminate lesion. The stakes of getting the call wrong are high. A false positive can lead to inappropriate upstaging, unnecessary systemic therapy, or even exclusion from potentially curative treatment. A false negative can delay therapy for genuinely metastatic disease. Equivocal findings also trigger extra imaging, prolonged follow-up, higher costs and considerable patient anxiety, and PSMA PET/CT findings have been shown to influence treatment decisions in up to 28 percent of cases.

Two competing strategies have emerged to tame this uncertainty. The first is PSMA-RADS Version 2.0, a five-point structured reporting scale that runs from certainly benign (score 1) to highly suspicious for malignancy (score 5), with the troublesome middle ground labelled score 3. Validation studies have reported excellent accuracy of around 95 to 97 percent across new diagnoses, follow-up and biochemical recurrence settings, along with strong agreement between different readers. The second is the BUMP score, a composite quantitative model that combines three readily obtainable variables: the maximum standardised uptake value (SUVmax) from the PET scan, the mean Hounsfield units (HU) measuring lesion density on CT, and whether the patient was on androgen deprivation therapy at the time of imaging. In its original development and external validation, BUMP achieved area under the receiver operating characteristic curve (AUROC) values of 0.871 and 0.925 respectively, raising hopes that a simple formula could standardise and even automate this difficult judgement.

To test that hope in the real world, researchers led by James Cairns and Salman Arain of Leeds Teaching Hospitals NHS Trust, working with colleagues in Bristol, Newcastle and Oxford, retrospectively analysed [18F]PSMA-1007 PET/CT scans performed as part of routine clinical care at three UK tertiary cancer centres between October 2021 and March 2023. The cohort comprised 130 patients with a median age of 70 years, in whom 412 bone lesions with focal PSMA uptake were identified and segmented by blinded, experienced reporters using a paired-review process. Ground truth for each lesion was established through a composite reference standard combining follow-up imaging, prostate-specific antigen (PSA) kinetics and multidisciplinary clinical consensus, with a median follow-up of 22.8 months.

The biological signal the model relies on was clearly present in the data. Malignant lesions, which made up 60.9 percent of the sample (251 of 412), showed markedly higher radiotracer avidity than benign ones, with a median SUVmax of 7.54 versus 3.53, and greater CT density, with a median of 259 Hounsfield units versus 196. Malignant lesions were also larger, at a median of 12 millimetres compared with 6 millimetres for benign findings. The ribs were the most common site of focal uptake (42.2 percent of lesions), followed by the vertebrae (27.4 percent) and pelvis (15.3 percent). Most patients, 85.4 percent, were scanned for suspected recurrence after treatment, and the majority were not on androgen deprivation therapy at the time of imaging.

Against this backdrop, the BUMP score performed respectably but not brilliantly. Its median value was significantly higher in malignant lesions (0.427) than benign ones (0.08), and ROC analysis yielded an AUROC of 0.812 (95 percent confidence interval 0.636 to 0.931), a modest drop from the 0.87 and 0.92 reported in the original development work, likely reflecting the broader heterogeneity of a real-world multicentre population. Notably, the optimal probability threshold in this cohort, derived using the Liu criterion, was 0.134, considerably lower than the 0.25 cut-off published in the literature. At the locally derived threshold, BUMP achieved a sensitivity of 73.9 percent and a specificity of 80.0 percent, with an F1 score of 0.800; at the literature threshold, sensitivity fell to 62.6 percent while specificity rose to 88.5 percent. The authors emphasise that this threshold shift is a common challenge when generalising quantitative models across institutions and underscores the need for local calibration before clinical deployment.

Human-led structured reporting, by contrast, was in a different league. PSMA-RADS v2.0 scoring achieved an AUROC of 0.980 (95 percent CI 0.952 to 0.993). When scores of 1 to 2 were treated as benign and 4 to 5 as malignant, with the 44 equivocal score-3 lesions excluded, the framework delivered a sensitivity of 98.3 percent and a specificity of 96.2 percent, for a balanced accuracy of 97.3 percent and an F1 score of 0.981. The difference in AUROC between PSMA-RADS and BUMP was 0.1679 (95 percent CI 0.050 to 0.337, p < 0.001). Sensitivity analyses confirmed the result was robust: whether score-3 lesions were counted as benign or malignant, and whether analysis was performed at the lesion or patient level, PSMA-RADS retained its clear advantage, with patient-level AUROCs of 0.966 versus 0.840.

Most telling of all was BUMP’s performance precisely where a quantitative adjunct was supposed to shine: the equivocal lesions. In the 44 lesions that experts had labelled PSMA-RADS 3, the BUMP score’s discriminative power collapsed to an AUROC of just 0.605 (95 percent CI 0.457 to 0.781), barely better than a coin flip. The reason, the authors argue, is that borderline lesions tend to show intermediate SUV values and ambiguous CT density, exactly the inputs BUMP depends on, while expert readers integrate information the formula cannot see: lesion distribution, morphology, anatomical location, overall disease burden and the patient’s pre-test probability. Illustrative cases drove the point home. A vertebral metastasis with a correlative CT abnormality was correctly called malignant by PSMA-RADS but scored benign (0.1) by BUMP, while a hot facet joint from benign arthropathy, with an SUVmax of 7.6, was correctly dismissed by the human reader but flagged as likely malignant (0.61) by the score.

The study’s conclusions come with caveats. The retrospective design introduces selection bias, follow-up duration varied widely, lesion-level statistics were not adjusted for intra-patient clustering, BUMP was not formally assessed for calibration as an absolute probability, and the findings are specific to [18F]PSMA-1007, which has a higher burden of unspecific bone uptake but a lower malignancy rate than other tracers. Ground truth itself relied on multidisciplinary consensus rather than histology, raising the possibility of incorporation bias. Still, the authors see a role for BUMP as a decision-support adjunct, particularly in high-volume centres or for reporters with limited subspecialist experience, and point to more sophisticated radiomics and artificial intelligence models, which incorporate contextual awareness, as the likely next frontier. For now, the message is clear: structured human expertise, in the form of PSMA-RADS v2.0, remains the gold standard for interpreting the skeleton’s ambiguous signals, and quantitative scores should inform, not replace, the expert eye.

Subject of Research: Comparison of the BUMP score and PSMA-RADS v2.0 for characterising focal bone uptake on PSMA-1007 PET/CT in prostate cancer

Article Title: Comparative analysis of PSMA-RADS Version 2.0 versus bone uptake metastatic probability (BUMP) score for characterising focal bone uptake on [18F]PSMA-1007 PET/CT in prostate cancer – a multicentre evaluation

Article References: Cairns, J., Arain, S., Challapalli, A., Stavropoulos, A., Ali, T., Ward, D., Sage, J., Ionescu, G., Willaime, J. M. Y., Frood, R., Bahl, A., Petrides, G., & Scarsbrook, A. (2026). Comparative analysis of PSMA-RADS Version 2.0 versus bone uptake metastatic probability (BUMP) score for characterising focal bone uptake on [18F]PSMA-1007 PET/CT in prostate cancer – a multicentre evaluation. European Journal of Nuclear Medicine and Molecular Imaging. https://doi.org/10.1007/s00259-026-08154-0

Image Credits: AI Generated

DOI: 10.1007/s00259-026-08154-0

Keywords: prostate cancer, PSMA PET/CT, PSMA-1007, bone metastases, BUMP score, PSMA-RADS, unspecific bone uptake, nuclear medicine, diagnostic accuracy, SUVmax, androgen deprivation therapy, structured reporting