Prostate cancer is the second most common cancer in men worldwide, with an estimated 1.4 million new cases and 375,000 deaths in 2020. However, a large proportion of diagnosed tumors are clinically insignificant -- slow-growing cancers that will never cause symptoms or harm during a patient's lifetime. Treating these unnecessarily leads to significant side effects without benefit.
Clinically significant prostate cancer (csPCa) is defined as having a Gleason score of 3+4 or higher (ISUP Grade 2 or above), indicating disease that may progress, metastasize, and cause death. Accurately distinguishing csPCa from insignificant tumors is critical to avoid overtreatment of the many while ensuring aggressive management of those who truly need it.
The standard tool for identifying suspicious lesions is multiparametric MRI (mpMRI), which is scored on the PI-RADS system (1-5 scale). However, PI-RADS has a significant limitation: even for lesions scored 4 or higher, the positive predictive value for csPCa is only about 49%. This means roughly half of patients recommended for biopsy based on MRI alone do not have significant cancer.
Prostate biopsy -- the gold standard for diagnosis -- is an invasive procedure associated with complications including bleeding, infection, pain, urinary symptoms, and erectile dysfunction. Reducing unnecessary biopsies while maintaining cancer detection is therefore a major clinical priority.
Radiomics is a field that extracts large numbers of quantitative features from medical images -- capturing texture, shape, intensity patterns, and spatial heterogeneity that are invisible to the naked eye. These features can serve as objective imaging biomarkers that reflect the underlying biology of tumors.
Most prior radiomics studies in prostate cancer focused exclusively on the intratumoral region -- the tumor itself. However, emerging research highlights the importance of the peritumoral region, the tissue immediately surrounding the tumor. Cancer cells release signals that reshape this neighborhood through hypoxia, new blood vessel formation, and immune cell infiltration, all of which may be detectable on imaging.
Magnetic resonance spectroscopy (MRS) is a technique that measures the metabolic profile of tissue, including levels of metabolites like choline (Cho), creatine (Cr), and citrate (Cit). In prostate cancer, citrate levels decrease while choline and creatine increase, and the elevated (Cho+Cr)/Cit ratio correlates with Gleason score. However, MRS has been little studied in combination with radiomics and clinical models.
A key barrier to adopting machine learning in clinical medicine is the black box problem: powerful models like gradient boosting make accurate predictions but cannot explain why. SHAP (SHapley Additive exPlanations) analysis, rooted in game theory, solves this by quantifying each feature's contribution to individual predictions, making AI models understandable to clinicians.
The study retrospectively analyzed 350 patients with suspicious prostate lesions who underwent 3.0 Tesla mpMRI before biopsy. Patients were divided into a training set (n=191), a testing set (n=83), and a temporal validation set (n=76, collected from a later time period) to rigorously assess the model's stability over time.
For each patient, an experienced radiologist manually drew two distinct regions of interest on both T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC) maps: the VOIintra (intratumoral region, tracing the tumor boundary) and the VOIperi (peritumoral region, extending 4mm outward from the tumor edge, with non-prostatic tissues such as the rectum and bladder removed).
The 4mm peritumoral margin was specifically chosen based on the tumor's small size and the 3.5mm slice thickness of the imaging protocol, ensuring enough surrounding tissue information was captured without including irrelevant structures. A second radiologist independently segmented 30 randomly selected cases to evaluate reproducibility.
Clinicopathological variables collected included age, PSA, PSA density (PSAD) (PSA divided by prostate volume), digital rectal exam (DRE) result, smoking status, tumor location (peripheral zone vs. transition zone), and PI-RADS score. Metabolic data (Cho, Cr, Cit levels and the (Cho+Cr)/Cit ratio) were obtained via MRS for each lesion.
From each region of interest, a total of 1,702 radiomic features per region were extracted using the pyradiomics library, spanning shape, first-order intensity statistics, texture matrices (GLCM, GLRLM, GLSZM, GLDM, NGTDM), and high-order wavelet-filtered features. This yielded over 6,800 candidate features per patient across both regions and both imaging types.
A rigorous multi-step feature selection pipeline was applied: features with poor inter-reader reproducibility (intraclass correlation below 0.8) were discarded, followed by statistical significance testing, Pearson correlation filtering to remove redundant features, and finally recursive feature elimination (RFE) with 5-fold cross-validation to identify the optimal feature subset. This yielded 10 features for the intratumoral score and 16 for the peritumoral score.
The final radiomic scores (intra-rad-score and peri-rad-score) were constructed as linear combinations of the selected features. Five prediction models were then built using the XGBoost algorithm: a clinical model, an MRS model, an intra-radiomics model, a peri-radiomics model, and a combined model integrating all feature types.
XGBoost (eXtreme Gradient Boosting) is an ensemble machine learning method that builds many decision trees sequentially, each correcting errors from the previous one. It was chosen for its strong predictive performance, resistance to overfitting, and compatibility with SHAP-based interpretability analysis.
The combined model achieved the best performance across all metrics. In the testing set, it reached an AUC of 0.968, accuracy of 92.8%, sensitivity of 93.2%, and specificity of 92.3% -- outperforming the clinical model alone (AUC 0.868) and the MRS model (AUC 0.824). In the temporal validation set, the combined model maintained an AUC of 0.940.
Both radiomic models -- intra-radiomics (AUC 0.936) and peri-radiomics (AUC 0.943) -- individually outperformed the clinical and MRS models in the testing set, and neither was statistically inferior to the combined model, indicating that radiomic features captured most of the diagnostic signal. Adding clinical and metabolic variables pushed performance further while also providing interpretability.
SHAP analysis of the combined model revealed that the most important predictors were, in order: intra-rad-score, PSAD, peri-rad-score, PI-RADS score, and PSA. Citrate (Cit) had only marginal importance and was the only metabolic feature contributing meaningfully. The MRS-derived (Cho+Cr)/Cit ratio contributed very little despite its known biological relevance.
In the SHAP visualizations, high values of the top five predictors consistently pushed predictions toward csPCa (positive influence), while high citrate levels pushed predictions away from csPCa -- consistent with the known biology where cancer tissue has reduced citrate production. Individual patient-level SHAP force plots allowed case-by-case review of why the model made each specific prediction.
The model's reliance on intratumoral radiomic features reflects the known histopathology of clinically significant prostate cancer: higher-grade tumors (Gleason 4 or 5) show increased cellular density, reduced extracellular space, poorly formed glands, cribriform patterns, and fused glandular structures -- all microscopic changes that alter the texture and intensity patterns captured by MRI-based radiomics.
The contribution of peritumoral radiomics reflects important biological processes in the tumor microenvironment. Rapidly growing cancer cells create local hypoxia in surrounding tissue, which triggers the release of hypoxia-inducible factors (HIF-1) that stimulate new blood vessel and lymphatic vessel formation. Additionally, immune cell infiltration increases the heterogeneity of the peritumoral zone -- all captured as texture changes on imaging.
PSAD emerged as the most important single clinical variable, consistent with previous research. PSA density normalizes PSA for prostate size, making it a more specific indicator of cancer than raw PSA. Multiple prior studies have shown PSAD as a key independent predictor of csPCa, particularly in PI-RADS 3 and 4 lesions.
The limited performance of MRS may reflect technical challenges rather than a true absence of metabolic signal: MRS has long scan times, low spatial resolution, and was removed as a primary examination method from PI-RADS v2.1 guidelines. The researchers note that newer metabolic imaging methods like chemical exchange saturation transfer (CEST) MRI and hyperpolarized carbon-13 MRI may offer more reliable metabolic information in future models.
The combined model's high specificity (92.3% in testing, 87.5% in validation) means it can correctly identify a large proportion of men with non-significant disease -- men who could safely avoid biopsy and its associated risks of infection, bleeding, pain, and urinary dysfunction. This is particularly valuable for the large pool of men with PI-RADS 3-4 lesions where the positive predictive value of MRI alone is poor.
The SHAP interpretability framework transforms what would be a black-box prediction into a transparent clinical tool. Clinicians can review which features drove the prediction for a specific patient and in which direction, supporting rather than replacing clinical judgment. This addresses a key barrier to AI adoption in medical practice.
The model's performance remained high on the temporal validation set despite some differences in patient demographics over time (notably a shift in tumor location toward the transition zone), suggesting reasonable stability and generalizability across modest population changes within the same institution.
This study developed the first comprehensive interpretable machine learning model to simultaneously integrate intratumoral radiomics, peritumoral radiomics, clinical data, MRI scores, and MR spectroscopy metabolic data for predicting clinically significant prostate cancer -- achieving excellent performance (AUC 0.968) validated across two independent test sets.
Key limitations include the single-center retrospective design, which limits generalizability. The peritumoral zone definition of 4mm is not standardized across the field, and future studies should systematically compare different peritumoral expansion distances. Manual tumor segmentation introduces variability despite inter-reader reproducibility analysis.
External validation across multiple institutions and different MRI scanner types will be essential before clinical deployment. Prospective studies could also assess whether model-guided biopsy decisions result in measurably fewer unnecessary procedures while maintaining detection rates for significant cancers.