A machine learning radiomics model integrating prostate-specific membrane antigen positron emission tomography/computed tomography and multiparametric magnetic resonance imaging showed promise for risk stratification among patients with suspected prostate cancer, according to a retrospective multicenter study published in BMC Medicine.
Researchers evaluated 488 patients who underwent both imaging modalities prior to biopsy confirmation between February 2019 and October 2025. The Xiangya Hospital cohort included 366 patients and was split into training and internal test cohorts; two external validation cohorts included 41 patients from Qilu Hospital and 81 patients from The First Affiliated Hospital of Guangzhou Medical University.
Clinically significant prostate cancer was defined as Gleason score 3 + 4 or higher. For the primary classification task, the comparator group included both benign prostatic disease and Gleason score 3 + 3 prostate cancer, a clinically heterogeneous group that may affect interpretation of model performance.
Radiomic features were extracted from diffusion-weighted imaging, T2-weighted imaging, positron emission tomography, and computed tomography. Six machine learning classifiers were evaluated: logistic regression, support vector machine, Random Forest, Extra Trees, XGBoost, and LightGBM.
For clinically significant prostate cancer, multimodal models generally outperformed single-modality approaches. In the internal test cohort, logistic regression, support vector machine, and LightGBM each achieved an area under the curve of 0.91. In external validation, LightGBM achieved an AUC of 0.82 in the 41-patient Qilu Hospital cohort and 0.89 in the 81-patient Guangzhou Medical University cohort; logistic regression achieved AUCs of 0.80 and 0.85, respectively. AUC is a measure of how well a model discriminates between groups, where 1.0 indicates perfect discrimination.
The model’s clinical role remains uncertain. Negative predictive values were modest across several analyses, limiting conclusions about whether the approach could safely help defer biopsy. The researchers also did not directly compare the radiomics models with standardized PI-RADS assessment, PSA density, clinical risk calculators, or physician interpretation.
The study also evaluated automated whole-gland segmentation with an nnU-Net model. Automated segmentation produced somewhat lower but broadly similar model performance compared with expert-defined volumes of interest. Approximately 80% of extracted radiomic features showed good agreement between expert and automated segmentation.
The automated segmentation analysis had some methodological ambiguity. The manuscript states that an intraclass correlation coefficient threshold of 0.75 or higher indicated good agreement for the reproducibility analysis, but a later methods section states that features with an ICC below 0.85 were excluded, without clearly distinguishing how those thresholds were applied. That ambiguity makes it difficult to determine exactly which features entered the final models.
The researchers noted that positron emission tomography-derived features contributed most strongly to model predictions, with diffusion-weighted imaging features adding complementary information. T2-weighted and computed tomography features contributed less prominently.
Limitations included the retrospective design, small external validation cohorts, possible selection bias, imaging protocol variability across centers, and lack of prospective validation. Generalizability was also limited because all cohorts came from Chinese academic centers and used gallium-68 PSMA-617; performance in community practice, broader patient populations, or with other PSMA tracers remains unknown.
The findings support further study of multimodal radiomics as a decision-support tool for prostate cancer risk stratification, rather than a replacement for histopathologic confirmation.
The researchers reported no competing interests. The study was supported by the Hunan Provincial Science Fund for Distinguished Young Scholars, the National Natural Science Foundation of China, and other institutional and clinical research grants.
Source: BMC Medicine