Prostate cancer is the most common cancer in men and the fifth leading cause of cancer-related deaths worldwide, with approximately 1.4 million new diagnoses and 375,000 deaths in 2020 alone. Because the disease often develops slowly in its early stages, older men may not notice symptoms until the cancer has advanced -- making reliable early detection tools critical for reducing mortality.
Current standard diagnostic methods each have significant limitations. Digital rectal examination (DRE) is subjective and varies between examiners. The PSA blood test lacks specificity, producing many false positives that lead to unnecessary biopsies. Transrectal ultrasound-guided biopsy -- still the definitive diagnostic step -- samples at least 12 sites from the prostate and carries risks of bleeding and infection, yet can still miss tumors through random sampling error.
The Gleason score (GS) and the newer ISUP Grade Group (GG) system are the primary tools for classifying tumor aggressiveness. Gleason grades describe how abnormal gland architecture looks under a microscope, from well-organized glands (grade 1) to completely disorganized infiltrating cells (grade 5). Grade Group 1 corresponds to Gleason 6 or below (lowest risk), while Grade Group 5 corresponds to Gleason 9-10 (highest risk). Accurate grading determines treatment selection but requires tissue biopsy.
Multiparametric MRI (mpMRI) -- combining T2-weighted anatomy, diffusion-weighted imaging, and dynamic contrast-enhanced sequences -- has emerged as a powerful noninvasive approach. When combined with radiomics (mathematical extraction of image features) and artificial intelligence, mpMRI offers a path toward reducing reliance on invasive procedures and enabling more personalized cancer management.
Multiparametric MRI is the combination of several complementary imaging sequences: T2-weighted imaging (T2WI) shows detailed anatomy including gland zones and lesion structure; diffusion-weighted imaging (DWI) measures water movement in tissue, with cancer cells restricting water diffusion due to their dense packing; and dynamic contrast-enhanced (DCE) imaging tracks how injected contrast agent flows into and out of tissue, reflecting blood vessel density typical of tumors. Together these sequences allow precise lesion localization, staging, and biopsy guidance.
The PI-RADS (Prostate Imaging Reporting and Data System) provides a standardized framework for interpreting and reporting prostate MRI findings. Updated versions from the first edition in 2012 through PI-RADS v2.1 in 2019 have progressively refined the criteria. Lesions are scored 1-5: scores of 1 or 2 suggest benign findings, a score of 3 is equivocal and may require biopsy based on clinical context, and scores of 4 or 5 indicate likely clinically significant cancer requiring biopsy.
Research consistently shows mpMRI outperforms standard systematic biopsy for detecting clinically significant cancer. MRI-targeted biopsy achieves higher detection rates for significant cancer while reducing diagnosis of insignificant tumors that might otherwise be overtreated. Combining MRI-targeted biopsy with systematic biopsy improves detection rates by an additional 10% compared to either approach alone.
Despite its value, mpMRI has important limitations: results vary substantially between radiologists, patients with pacemakers or metal implants cannot safely undergo MRI, and a negative mpMRI result does not completely exclude cancer -- particularly in younger patients. These limitations motivate the development of AI-assisted tools to standardize interpretation and quantify risk more objectively.
The radiomics process follows a standardized four-step pipeline. Image acquisition captures mpMRI using standardized protocols; initiatives like the Quantitative Imaging Biomarkers Alliance (QIBA) and International Biomarker Standardization Initiative (IBSI) promote consistency across scanners. Preprocessing steps including denoising, standardization, and resampling ensure image data is uniform before analysis.
Segmentation -- delineating the tumor boundaries on each MRI slice -- is the most labor-intensive step. Manual segmentation by radiologists is time-consuming and subject to both intraobserver (same person, different times) and interobserver (different people) variability. Semi-automatic tools like 3D-Slicer and ITKSNAP help clinicians trace regions more efficiently, while fully automatic segmentation using convolutional neural networks (CNNs) is advancing rapidly. One CNN-based system achieved Dice similarity coefficients of 0.70 and sensitivity of 0.87 for automatic prostate lesion segmentation.
Feature extraction converts the segmented tumor region into hundreds of numerical measurements. These include shape features (size, roundness), first-order statistical features (brightness distribution), and texture features captured by matrices including GLCM (gray-level co-occurrence matrix), GLSZM (gray-level size zone matrix), and GLRLM (gray-level run length matrix). Deep learning also contributes deep radiomic features extracted from CNN intermediate layers, which capture patterns too subtle for human observation.
Feature selection reduces the high-dimensional feature space to prevent overfitting. Common methods include LASSO (which forces most feature weights toward zero), Random Forest importance scoring, and dimensionality reduction via Principal Component Analysis. Table 2 in the review paper summarizes results from multiple studies using these techniques, consistently showing that reduced, carefully selected feature sets outperform full feature sets for predicting clinical outcomes like Gleason score and clinically significant cancer status.
Radiomics models for prostate cancer have shown substantial performance across multiple clinical tasks. For Gleason score prediction, a texture feature model using T2WI and ADC maps with random forest classification achieved AUCs of 83.4%, 72.7%, and 77.4% for three Gleason score groups. A joint intensity matrix feature approach achieved AUC 78.4% for low-grade, 82.4% for intermediate-grade, and 64.8% for high-grade cancer. When deep CNN-extracted features replaced handcrafted features, performance improved substantially.
For lymph node invasion prediction, a radiomic model using SVM achieved AUC of 0.915 on the test set, suggesting the potential to avoid extended pelvic lymph node dissection in many patients. A radiomics-based risk nomogram showed it could avoid up to 80% of extended dissections while missing only 1.1% of patients with true lymph node involvement. For Gleason score improvement over PI-RADS alone, combining a radiomics score with PI-RADS v2.1 achieved AUC 0.861 versus 0.845 for PI-RADS alone.
Despite these promising results, model stability is a major unresolved challenge. Radiomic features extracted from the same image can vary substantially based on scanner type, imaging protocol, field strength, and even the specific software used for extraction. One study found that only 25 of 1,409 radiomics features showed high robustness across different scanner protocols. Another found that the percentage of stable features across normalization methods ranged from only 3.4% to 8%. This means the majority of features most studies use may not be reproducible across sites.
Generalizability is equally challenging. When models developed at single institutions are tested at different institutions using different equipment and protocols, performance often drops substantially. One study found external validation AUC of 0.54 versus internal cross-validation AUC of 0.75 -- a 0.21 point drop. Federated learning -- training models across multiple institutions without sharing raw patient data -- is an emerging approach to building models that generalize while maintaining patient privacy.
Radiogenomics combines imaging-derived radiomic features with genomic data, creating models that can noninvasively estimate genetic tumor characteristics. For prostate cancer, this is particularly valuable because genomic testing -- identifying mutations in genes like BRCA1/2, ATM, and ERG -- currently requires expensive, time-consuming tissue analysis but guides critical treatment decisions including eligibility for PARP inhibitors and immunotherapy.
Key genetic associations have been identified through radiogenomics research. PTEN loss and ERG-positive expression correlate with MRI lesion visibility -- MRI-invisible lesions show less PTEN loss than visible ones. BRCA2 carriers have higher prostate cancer incidence and faster progression, and their tumors often show distinctive aggressive MRI features before treatment begins. Eleven microRNAs have been identified as sensitive early detection biomarkers, with ADC values correlating with specific miRNA expression patterns.
Hypoxia-related genes including ANGPTL4, VEGFA, and P4HA1 show correlation with MRI texture features -- meaning radiomic texture measurements may indirectly reflect the oxygen status of the tumor, which affects treatment resistance. A hypoxia gene signature called Ragnum-signature, derived from biopsy tissue, predicts prognosis and can be correlated with mpMRI features, enabling non-invasive hypoxia assessment from imaging alone.
A key limitation of radiogenomics is data heterogeneity: radiomic measurements and genomic measurements come from different platforms with different sources of variability, and linking them requires large matched datasets that are difficult and expensive to collect. The ImaGene web platform is one emerging tool that helps researchers build AI models linking oncology imaging data with gene expression databases, but multicenter standardization remains an open challenge.
Multi-omics refers to the simultaneous integration of data from multiple molecular measurement platforms: genomics (DNA sequences and mutations), transcriptomics (RNA gene expression), proteomics (protein expression and modifications), epigenomics (DNA methylation and histone modifications that regulate gene expression without changing DNA sequence), and metabolomics (small molecule metabolites reflecting active cellular chemistry).
Each omics layer provides a distinct window into cancer biology. Transcriptomics of prostate cancer cell lines has identified specific gene expression signatures of treatment resistance. Single-cell sequencing has identified a population of luminal progenitor cells as possible contributors to prostate carcinogenesis. Metabolomics has identified 26 metabolites significantly altered in prostate cancer tissue, pointing to dysregulation in amino acid, purine, and glycerophospholipid metabolism pathways as active drivers of tumor growth.
Multi-omics models combining several data types consistently outperform single-omics models. Studies integrating genomics, methylomics, and transcriptomics achieve better risk stratification than any single approach, identifying multi-omics signatures independently predictive of disease relapse. Combined analysis of mRNA, microRNA, DNA methylation, and copy number variations identifies molecular clusters with distinct clinical behavior that single-gene tests miss.
Integrating radiomics with multi-omics -- creating truly comprehensive models that span from MRI imaging through molecular biology -- remains an early-stage research direction with significant promise. Few existing studies have combined all layers, and those that have are limited by small sample sizes and non-standardized methods. Large collaborative database initiatives are needed to enable the training of robust integrated models.
The majority of current radiomics studies in prostate cancer share common structural limitations: they are single-center, retrospective, and have small sample sizes. These features inflate apparent model performance and limit generalizability to other clinical settings. Moving the field forward requires multicenter, prospective studies with hundreds or thousands of patients across diverse imaging equipment and demographics.
Technical gaps persist across the radiomics pipeline. Most studies do not include DCE-MRI sequences despite their routine collection in clinical prostate MRI, leaving potentially valuable information unused. Automatic segmentation algorithms need further improvement to reliably delineate tumor boundaries without expert manual tracing. Inaccurate biopsy-derived ground truth labels -- which arise from sampling errors in needle biopsy -- directly limit the accuracy of radiomics models trained on them.
The gap between AI model performance in research settings and actual clinical implementation remains wide. Bridging it requires solving the interpretability problem -- clinicians need to understand why a model predicts high risk, not just that it does. It also requires standardizing evaluation criteria, completing multi-reader studies comparing AI to radiologist performance, and conducting prospective evaluation in real clinical workflows.
Despite these challenges, AI-enhanced radiomics is considered a highly promising clinical tool. When combined with radiogenomics and multi-omics data, it offers a vision of comprehensive, noninvasive prostate cancer profiling from blood tests and MRI alone -- reducing the need for painful biopsies, providing earlier and more accurate risk stratification, and enabling treatment decisions tailored to each patient's individual tumor biology.