Significant prostate cancer prediction may become more precise when MRI-based radiomics and clinical information are combined with PI-RADS assessment. A 2026 publication in Insights into Imaging evaluated whether a radiomics-based model could help distinguish significant prostate cancer, compare with PI-RADS and support biopsy decision-making. The analysis used MRI cases from men with suspected prostate cancer at Vall Hebron Hospital in Barcelona. Biopsy samples provided the reference standard, with significant disease defined by Gleason Grade 7 or higher. The strongest results came from a combined model that brought together PI-RADS, radiomic features and selected clinical variables. Radiomics alone performed similarly to PI-RADS, while the broader combined approach improved discrimination and showed potential to reduce false positives and unnecessary biopsies.
Clinical Context and Model Design
The work focused on men with suspected prostate cancer based on raised prostate-specific antigen and/or abnormal digital rectal examination. Eligible cases included prostate MRI interpreted with PI-RADS and biopsy performed within three months. Men with a previous prostate cancer diagnosis, active surveillance, missing data, biopsy before MRI or image artefacts that prevented accurate reading were excluded.
The final cohort included 1497 MRI cases from 1395 men. Cases were randomly divided into training and test groups. Significant prostate cancer was present in just under two fifths of the overall cohort. Most men had no previous prostate biopsy, while a smaller group had a previous negative biopsy. The training and test groups were similar across the assessed clinical characteristics, including cancer detection rate and PI-RADS score.
Four predictive models were developed. One used PI-RADS alone. One used radiomics alone. One combined PI-RADS and radiomics. The final model combined PI-RADS, radiomics and clinical variables. The clinical variables included age, prostate-specific antigen, prostate volume, digital rectal examination, previous negative biopsy, family history and PI-RADS index lesion score.
Combining Imaging and Clinical Variables
Radiomic features were extracted from automatically segmented prostate gland images. The process used prostate MRI sequences and generated imaging maps. Seminal vesicle masks were not used because prostate carcinoma does not originate in that region. The modelling process tested several classifiers, with Random Forest giving the best overall performance among those assessed.
The PI-RADS model and the radiomics model achieved comparable discrimination in the validation cohort. Radiomics alone did not significantly outperform PI-RADS. Combining PI-RADS with radiomics improved performance, but the improvement was not statistically significant when compared with the standalone models.
Must Read: Accelerating Prostate MRI With AI Reconstruction
The best discrimination came from the model that combined radiomics, PI-RADS and clinical variables. Its area under the curve reached 0.891, which was significantly higher than the other three models. The result indicates that clinical information added value when paired with imaging assessment and radiomic features. PI-RADS remained a central component of the strongest model, rather than being replaced by radiomics.
Model explainability showed that PI-RADS score had the greatest influence on prediction, followed by age. Higher values for these factors were linked with greater predicted risk of significant prostate cancer. Larger prostate volume was linked with a lower predicted chance of significant disease. Radiomic features showed variable effects, with features derived from diffusion-related imaging contributing more consistently to predictions.
Biopsy Avoidance and Clinical Utility
Clinical utility was assessed by fixing sensitivity to match the PI-RADS threshold used for biopsy decisions. This allowed comparison of specificity and biopsy avoidance across the models while keeping the undetected significant cancer rate unchanged. At this setting, the PI-RADS model avoided some biopsies while maintaining high sensitivity. Radiomics alone avoided slightly fewer biopsies.
The combined PI-RADS and radiomics model improved specificity and biopsy avoidance compared with the standalone models. Adding clinical variables produced the highest specificity and the highest biopsy avoidance rate, reaching 18.15%. However, the differences in biopsy avoidance and specificity did not reach statistical significance. The undetected significant prostate cancer rate remained the same across all four models because sensitivity was fixed.
Decision curve analysis also favoured the combined approaches across a broad range of threshold probabilities. The models that included both PI-RADS and radiomics, with or without clinical variables, showed higher net clinical benefit than PI-RADS alone or radiomics alone. The clinical utility results showed that combined models allowed more biopsies to be avoided while maintaining similar rates of undetected significant prostate cancer.
Several limitations remain important. The cohort came from a single centre and a single MRI vendor, which may affect generalisability. PI-RADS scoring used two versions of the system. Deep radiomics and lesion-specific feature extraction were not tested. External validation was not performed because the required clinical variables were difficult to obtain in external cohorts.
The combined use of radiomics, PI-RADS and clinical variables improves discrimination of significant prostate cancer in men with suspected disease. Radiomics alone performs similarly to PI-RADS, but the strongest model comes from integrating imaging assessment, radiomic features and clinical information. The combined approach also shows the highest biopsy avoidance rate when sensitivity matches the PI-RADS biopsy threshold, although this clinical efficacy difference is not statistically significant. Prospective validation and further methodological refinement remain necessary before routine clinical integration.
Source: Insights into Imaging
Image Credit: iStock
References:
Antolin A, Mast R, Roson N et al. (2026) Combined radiomics, PI-RADS, and clinical model improve significant prostate cancer prediction and guide biopsy decision. Insights Imaging; 17, 118.
