Multiparametric MRI is the standard imaging approach for detecting and characterizing prostate cancer, but its interpretation relies heavily on the subjective visual assessment of radiologists. The PI-RADS (Prostate Imaging Reporting and Data System) provides a standardized scoring framework from 1 to 5, but even with this system, diagnostic accuracy varies considerably between readers and institutions.
Radiomics is an approach that extracts hundreds or thousands of quantitative features from medical images, including measures of texture, intensity distribution, shape, and spatial relationships. These features can capture image information that is not perceptible to the human eye, potentially improving diagnostic performance beyond what visual assessment alone can achieve.
A key clinical challenge in prostate cancer is distinguishing clinically significant prostate cancer (csPCa) from clinically insignificant prostate cancer (cisPCa). Only clinically significant cancers require active treatment; insignificant cancers can be managed with active surveillance. An imaging biomarker that reliably makes this distinction could reduce unnecessary biopsies and overtreatment.
The study enrolled 191 patients who underwent multiparametric MRI followed by biopsy. Radiomics features were extracted from two MRI sequences: apparent diffusion coefficient (ADC) maps derived from diffusion-weighted imaging and T2-weighted (T2w) images. These two sequences together provide complementary structural and functional information about lesion tissue.
A radiomics feature extraction pipeline computed a large number of quantitative features from manually segmented lesion regions, including first-order statistics, texture features from the gray-level co-occurrence matrix, and shape features. Feature selection was applied to identify the subset most predictive of cancer malignancy and significance while reducing overfitting risk.
The final model was an ensemble that combined the radiomics score with three established clinical parameters: PI-RADS score (the radiologist's visual assessment), PSA density (PSAD) (PSA level divided by prostate volume, a marker that corrects for prostate size), and digital rectal examination (DRE) result. The ensemble approach was designed to harness the complementary information from each input rather than relying on any single marker.
The dataset was split into training and test sets. Performance was evaluated separately for the tasks of distinguishing malignant versus benign lesions and distinguishing clinically significant versus clinically insignificant prostate cancer, using AUC as the primary metric.
The apparent diffusion coefficient (ADC) map is a quantitative map derived from diffusion-weighted MRI that reflects how freely water molecules move through tissue. Cancer tissue, with its densely packed cells, restricts water diffusion and shows low ADC values. Radiomics features extracted from ADC maps therefore capture not just mean diffusion values (which clinical readers already assess) but also spatial heterogeneity, texture, and the distribution of low-ADC voxels within a lesion.
Radiomics from T2-weighted images captures the morphological and textural properties of the lesion as seen on this anatomical sequence. Features such as intensity variance, entropy, and edge sharpness can reflect tumor architecture in ways that complement the functional information from ADC. Combining features from both sequences addresses the complementary nature of structural and functional MRI information.
PSA density adjusts the raw PSA blood test value by dividing by the prostate's total volume measured on MRI. A high PSA density suggests that the elevated PSA is not simply due to a large benign prostate but reflects cancer-related PSA production. This correction makes PSAD a more specific marker for clinically significant cancer than PSA alone.
In the test set, the ensemble model achieved an AUC of 0.889 for distinguishing malignant from benign lesions and an AUC of 0.844 for distinguishing clinically significant from clinically insignificant prostate cancer. These values represent strong diagnostic performance that exceeds what any single input achieves on its own.
For the clinically important csPCa versus cisPCa task, the ensemble model was significantly better than mean ADC alone (p equals 0.022). Mean ADC is the simple single-number measure of diffusion that radiologists routinely use; the finding that the ensemble significantly outperforms it demonstrates that the additional complexity of radiomics and clinical parameter integration is justified.
PI-RADS score alone also showed strong performance, as expected for an established and validated scoring system. The additional gain from combining PI-RADS with radiomics and clinical parameters demonstrates that quantitative imaging features capture information independent of what the radiologist extracts through visual assessment.
A central clinical motivation for this work is reducing the rate of unnecessary prostate biopsies. Current practice biopsies many men with elevated PSA or suspicious MRI findings, but a substantial proportion of these biopsies find no cancer or only indolent cancer that would never cause harm. Biopsy carries risks of bleeding, infection, and pain.
A tool with AUC greater than 0.88 for malignancy detection could be used to define risk thresholds that spare low-risk patients from biopsy while maintaining high sensitivity for clinically significant cancer. If the model identifies a patient as very low risk, shared decision-making could support deferring biopsy with close follow-up rather than immediate tissue sampling.
Integrating PSA density and DRE alongside imaging-derived scores also reflects a move toward holistic risk calculators rather than single-test decision-making. These clinical variables are routinely collected without additional cost or patient burden, making their inclusion in a multivariate model practically straightforward once the imaging analysis infrastructure is established.
This study establishes that combining radiomics features from ADC maps and T2w images with PI-RADS score, PSA density, and DRE in an ensemble model achieves better discrimination of clinically significant prostate cancer than any single marker, including PI-RADS alone or mean ADC alone.
The results reinforce a broader principle in clinical decision support: clinically complementary information sources, when properly integrated using machine learning, outperform any individual test. The challenge is developing validated, reproducible feature extraction and model implementation pipelines that can be deployed consistently across different institutions and scanner types.
Future work should focus on prospective validation across multiple institutions, testing whether the ensemble model's performance is maintained on images from different MRI vendors and field strengths, and on determining optimal decision thresholds that balance sensitivity and specificity in specific clinical contexts such as screening versus staging.