PET/MR Scan Predicts Prostate Cancer Relapse Before Surgery

For the hundreds of thousands of men who undergo radical prostatectomy each year, one question looms over every postoperative follow-up visit: will the cancer come back? A multicentre study published in the European Journal of Nuclear Medicine and Molecular Imaging now offers a strikingly precise answer to that question before the operation even begins. A team led by researchers at Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, has developed and externally validated a prognostic model built entirely from preoperative 18F-PSMA-1007 PET/MR imaging, showing that the combined molecular and anatomical picture of a patient’s tumour can predict biochemical recurrence up to three years after surgery. The work, registered retrospectively on ClinicalTrials.gov as NCT06604377, involved 210 patients across three centres and represents one of the most rigorous attempts yet to fuse hybrid imaging data with classical statistical modelling in prostate cancer.

The clinical problem the researchers set out to solve is deceptively simple to state but notoriously difficult to quantify. After surgical removal of the prostate, the earliest sign that cancer has returned is a rise in prostate-specific antigen, or PSA, the blood protein that gives biochemical recurrence its name. Roughly one in three patients in the study cohort eventually experienced this endpoint. But PSA alone is a lagging indicator: by the time it climbs, micrometastatic disease may already be established, and the window for effective salvage radiotherapy or androgen deprivation may have narrowed. Existing tools such as the CAPRA-S score, which combines pathological findings from the surgical specimen with PSA and Gleason grade, are useful but fundamentally retrospective, because they depend on tissue that can only be examined after the operation. A genuinely preoperative predictor would allow surgeons and radiation oncologists to plan adjuvant treatment strategies while the patient is still on the operating table.

The technology at the heart of the study is a hybrid imaging platform that simultaneously captures two complementary views of the same tumour. The PET component relies on 18F-PSMA-1007, a fluorine-18 labelled radioligand that binds to prostate-specific membrane antigen, a transmembrane enzyme overexpressed on the surface of nearly all prostate cancer cells. Because the tracer carries a positron-emitting isotope, its accumulation within a lesion is directly proportional to PSMA expression, which in turn correlates with tumour aggressiveness, hypoxia, and metastatic potential. The MR component provides high-resolution soft-tissue contrast, allowing precise delineation of tumour boundaries within the prostate and assessment of extracapsular extension. Performing both acquisitions in a single session eliminates the registration errors that plague separately acquired scans and yields co-registered datasets in which every voxel carries both metabolic and morphological information.

From these co-registered datasets the researchers extracted quantitative imaging features, a practice known as radiomics. Two derived parameters proved especially powerful. The first, ADCmin-TBR, combines the minimum apparent diffusion coefficient measured on diffusion-weighted MRI with the tumour-to-background ratio of PET tracer uptake. ADC reflects how freely water molecules diffuse through tissue: densely packed, aggressively proliferating tumour cells restrict water movement, driving ADC values down. TBR, by contrast, captures the metabolic intensity of the lesion relative to surrounding tissue. By multiplying or otherwise combining these two quantities, ADCmin-TBR effectively encodes a lesion that is both densely cellular and metabolically voracious, the hallmark of a tumour primed to recur. The second parameter, TMR, integrates tumour volume characteristics with metabolic activity, capturing the overall disease burden visible on the hybrid scan.

To build the model, the team divided their 210 patients into a training cohort of 150 and an external validation cohort of 60 drawn from different centres. Rather than hand-picking candidate predictors, they employed least absolute shrinkage and selection operator regression, or LASSO, coupled to a Cox proportional hazards framework. LASSO works by penalising model complexity, shrinking the coefficients of weak predictors toward zero and effectively eliminating them, which guards against overfitting in datasets where the number of variables approaches the number of events. The researchers used ten-fold cross-validation with the one-standard-error rule, a conservative criterion that selects the simplest model within one standard error of the best-performing one. Only three predictors survived this stringent selection: ADCmin-TBR, TMR, and the PET/MR-derived T-stage, an imaging-based assessment of how far the tumour has extended beyond the prostate capsule.

The performance figures are impressive for a purely preoperative tool. Time-dependent areas under the receiver operating characteristic curve, which measure discrimination at specific follow-up horizons, ranged from 0.703 to 0.858 in the training cohort and from 0.766 to 0.846 in the external validation cohort across the one-to-three-year window. An AUC of 0.85 at three years means the model correctly ranks a randomly chosen recurring patient above a randomly chosen non-recurring patient in 85 of 100 such pairs. Calibration plots confirmed that predicted probabilities matched observed recurrence rates, and inverse probability of censoring weighted Brier scores accounted for patients whose follow-up ended before the endpoint could be observed. Decision-curve analysis, which quantifies the net clinical benefit of acting on a model’s predictions across a range of risk thresholds, showed positive net benefit across most clinically relevant thresholds, suggesting the model could genuinely inform treatment decisions rather than merely describe risk.

Perhaps the most consequential finding concerns how the PET/MR model relates to the established CAPRA-S score. On its own, the imaging model did not significantly outperform CAPRA-S in raw discrimination, as DeLong tests revealed no statistically significant differences in AUC. But discrimination is not the only measure of a prognostic tool’s worth. When the researchers formally tested whether adding the PET/MR model to CAPRA-S improved the statistical fit of the combined model, the answer was a clear yes. Likelihood-ratio tests showed significant improvement at both two and three years in the derivation cohort, with p-values below 0.001, and at three years in the external validation cohort, with a p-value of 0.031. In practical terms, this means the imaging model captures prognostic information that the pathological score cannot, offering genuinely incremental value precisely at the longer follow-up horizons that matter most for treatment planning.

The team also confronted the messy realities of real-world oncology data through sensitivity analyses. Two potential confounders loomed large: persistent postoperative PSA, which may reflect incomplete tumour resection rather than true recurrence, and secondary treatments such as salvage radiotherapy initiated before biochemical recurrence was documented, which can artificially delay the observed endpoint. Reassuringly, the two imaging-derived predictors, ADCmin-TBR and TMR, remained directionally stable across all sensitivity analyses, suggesting their prognostic signal is robust to these distortions. The PET/MR T-stage effect, however, was attenuated after treatment censoring, a finding the authors interpret as evidence that imaging-based staging may partly capture disease biology that secondary treatment subsequently modifies. This nuance underscores why the researchers caution that prospective validation is required before the model can guide secondary-treatment decisions in routine practice.

The implications of the study extend well beyond the statistics. If a preoperative PET/MR scan can identify patients at high risk of recurrence, surgeons could tailor the extent of lymph node dissection, radiation oncologists could escalate or de-escalate adjuvant therapy, and patients could be stratified into trials of intensified systemic treatment. The fluorine-18 label of PSMA-1007 also carries practical advantages over gallium-68 based tracers, including a longer half-life of roughly 110 minutes that permits central radiopharmacy production and wider distribution, potentially making the approach scalable beyond specialised academic centres. Previous work has already established that high PSMA expression correlates with early PSA recurrence and that PET-based tumour segmentation predicts survival in the recurrent setting, but this study is among the first to combine metabolic, diffusion, and staging information from a single hybrid examination into an externally validated preoperative model.

Challenges remain before the model reaches the clinic. The retrospective design, the modest validation cohort of 60 patients, and the predominance of Chinese centres all warrant broader prospective testing across diverse populations. Radiomic features are also sensitive to scanner vendor, acquisition protocol, and reconstruction parameters, so multicentre harmonisation will be essential. Yet the direction of travel is unmistakable. Medicine is steadily moving from reactive diagnosis toward predictive, imaging-driven risk stratification, and this study demonstrates that a single preoperative hybrid scan can carry enough quantitative information to forecast the trajectory of prostate cancer years before PSA betrays the first sign of relapse. For patients facing the anxious arithmetic of post-surgical surveillance, that prospect transforms a routine PET/MR examination into something far more powerful: a crystal ball built from physics, chemistry, and statistics, refined by data from 210 patients and validated across three hospitals.

Subject of Research: Preoperative 18F-PSMA-1007 PET/MR prognostic modelling for predicting biochemical recurrence after radical prostatectomy in prostate cancer

Article Title: Development and validation of a preoperative 18F-PSMA-1007 PET/MR prognostic model for predicting biochemical recurrence in prostate cancer: a multicentre study

Article References: Hu, J., Chunyu, H., Yang, Y., Song, Y., Cheng, C., Tang, H., Chen, X., Cheng, R., Dai, J., Meng, H., Zuo, C., Zhao, J., & Li, B. (2026). Development and validation of a preoperative 18F-PSMA-1007 PET/MR prognostic model for predicting biochemical recurrence in prostate cancer: a multicentre study. European Journal of Nuclear Medicine and Molecular Imaging. https://doi.org/10.1007/s00259-026-08179-5

Image Credits: AI Generated

DOI: 10.1007/s00259-026-08179-5

Keywords: prostate cancer, PSMA PET, 18F-PSMA-1007, PET/MR, biochemical recurrence, radical prostatectomy, prognostic model, radiomics, CAPRA-S, LASSO-Cox regression, multicentre study, molecular imaging