Prostate cancer has the highest incidence in Northern Europe, Australia/New Zealand, the Caribbean, and North America, and is the most commonly diagnosed male cancer in roughly two-thirds of countries. In the U.S. alone, prostate, lung/bronchus, and colorectal cancers account for about 48% of new male cases, with prostate cancer representing 29% of 2024 diagnoses and the second leading cause of cancer death in men.
Current diagnosis depends on core needle biopsy (CNB) triggered by elevated prostate-specific antigen (PSA) levels or abnormal digital rectal examination (DRE). The subsequent morphological examination by a pathologist is demanding work: identifying subtle changes in glandular architecture and cellular atypia across multiple cores requires sustained concentration, and the complexity varies with biopsy modality and the number of cores obtained. AI enters the picture because it can extract quantitative, high-dimensional features from digital slides that go beyond what the human eye can consistently detect.
Multiple AI tools have already progressed beyond research into clinical deployment, with some receiving FDA approval for histopathology assessment, diagnostic imaging interpretation, and risk stratification. The research focus has expanded into radiomics (extracting mathematical features from images), pathomics (the equivalent for pathology slides), and treatment outcome prediction.
The authors conducted a narrative review searching MEDLINE/PubMed, Embase, Web of Science, IEEE Xplore, arXiv, and ClinicalTrials.gov from January 2019 to October 2025. Search queries combined prostate cancer terms with AI, machine learning, radiomics, pathomics, PSA, MRI, TRUS, PSMA PET/CT, biopsy, radiotherapy planning, and focal therapies.
They included prospective and retrospective studies, multicenter evaluations, systematic reviews, and guidelines that evaluated AI applied to prostate cancer screening, risk stratification, imaging, histopathology, biomarker discovery, or treatment planning. Single-case reports, editorials, non-prostate work, and purely technical papers without clinical validation were excluded. Preprints were included when methods were transparent.
For quality control, they assessed risk of bias using QUADAS-2 (for diagnostic accuracy) and PROBAST (for prognostic models), with adherence checks against TRIPOD, CONSORT-AI, and DECIDE-AI reporting guidelines. Because of the heterogeneity in populations, modalities, and reporting, they used narrative synthesis following SWiM (Synthesis Without Meta-analysis) guidance rather than pooling results statistically, emphasizing model validity, calibration, clinically meaningful thresholds, and robust external validation.
Performance metrics extracted included AUC (area under the receiver operating characteristic curve), sensitivity/specificity, positive/negative predictive values, Dice coefficient for segmentation tasks, and where available, decision-curve analysis and time-to-event outcomes via Kaplan-Meier and Cox proportional-hazards methods.
Traditional PSA-driven screening has a well-documented problem: only about 25% of men biopsied for elevated PSA actually have cancer. This leads to massive overbiopsy and overdiagnosis. PSA specificity can be as low as 6%, meaning the test frequently flags healthy men. AI aims to fix this by adding intelligence on top of existing screening modalities.
Micro-ultrasound: AI applied to micro-ultrasound (micro-US) imaging achieved an AUROC of 0.871 versus 0.753 for traditional clinical prediction models, with a major improvement in specificity (68% vs. 27%) at comparable sensitivity. This means the AI caught just as many real cancers while dramatically reducing false alarms.
Multimodal MRI-TRUS fusion: A multi-center study showed that a multimodal AI system combining MRI and transrectal ultrasound (TRUS) data outperformed radiologist MRI interpretation. The AI achieved 88% specificity compared to 78% for radiologists, with equivalent cancer detection rates. This cross-modality approach extracts complementary tissue information that neither modality captures alone.
Liquid biopsy and fragmentomics: An emerging non-invasive approach combines cell-free DNA (cfDNA) fragmentomics with circulating tumor DNA (ctDNA) and cell-free mRNA, all interpreted by AI. Fragmentomics analyzes the size and breakage patterns of DNA fragments circulating in the blood, which differ between cancer patients and healthy individuals. A 25-gene blood-based test (GeneVerify) demonstrated 90% sensitivity and 91% specificity, showing clinical feasibility as a non-invasive risk stratification tool.
MRI performance: In large reader studies, AI systems for prostate MRI achieved AUC of 0.91 compared to 0.86 for expert radiologists. A commercial AI tool evaluated on over 10,000 scans showed 50% fewer false positives and detected 20% fewer clinically insignificant cancers, meaning better discrimination between dangerous and harmless findings.
PI-RADS 3 lesions (equivocal cases): These borderline lesions are a critical diagnostic challenge. The review found that combining AI analysis with PSA density achieved 77.8% sensitivity and 93.1% NPV (negative predictive value) for clinically significant cancer. This could eliminate biopsies for 83.3% of patients with equivocal lesions while maintaining high cancer detection.
Tumor volumetrics: AI-derived intraprostatic tumor volume (V_AI) on multiparametric MRI provided independent prognostic value beyond standard clinical and radiologic factors. For radical prostatectomy patients, five-year metastasis AUC was 0.89 for V_AI vs. 0.79 for NCCN risk classification. In the radiation therapy cohort, seven-year metastasis AUC was 0.84 vs. 0.74 (P = .02), a statistically significant improvement in predicting which patients would develop metastatic disease.
Advanced architectures: Deep learning on multiparametric MRI using a ResNet50 feature extractor with multi-head attention achieved AUC 0.89 (PR-AUC 0.91) by fusing T2-weighted, diffusion-weighted, and dynamic contrast-enhanced sequences. Radiomics-based random-forest classifiers trained on second-order texture features achieved AUC 0.87 for overall risk and AUC 0.89 for high-risk group identification.
PSMA PET/CT: For prostate-specific membrane antigen PET/CT scans, a fully automated algorithm achieved sensitivities of 85% for primary prostate tumors, 91% for lymph node metastases, and 61% for bone metastases, with strong correlation to manual quantification (r = 0.62-0.96 for total lesion volume/uptake). The aPROMISE platform supports fully automated longitudinal lesion tracking for treatment-response assessment.
Automated cancer detection and grading: AI models on whole-slide images (WSI) now outperform pathologists in detecting subtle cancer regions. In a three-center real-world deployment, a clinical-grade system achieved sensitivity of ~0.99 and specificity of ~0.93 for cancer detection on biopsy slides. A comprehensive systematic review confirmed good-to-excellent diagnostic performance, with multiple studies reporting accuracies above 90%.
Workflow impact: A prospective study of AI-assisted pathology in routine clinical practice showed concrete efficiency gains: ~20% reduction in slide reading time, ~20% fewer immunohistochemistry (IHC) staining orders, and ~40% fewer second-opinion requests, all without loss of diagnostic accuracy. In an active-surveillance cohort, an AI detection algorithm achieved 0.96 sensitivity and 0.73 specificity, suggesting large proportions of benign slides could be safely auto-screened.
Integrative prognostic models: Advanced machine-learning pipelines combining H&E-stained histology with Ki-67 immunohistochemistry and clinicopathologic data outperformed traditional risk tools (CAPRA-S and Gleason-based approaches) for predicting biochemical recurrence after radical prostatectomy. The system reclassified patients more accurately across risk groups by leveraging AI-driven quantification of IHC proliferation signals alongside standard clinical variables.
Aggressive phenotype stratification: Emerging evidence shows that AI-assisted analyses fusing morphometric features from histology with genomic/biomarker data can identify patients at higher risk of neuroendocrine differentiation or castration-resistant progression. These signals may inform more personalized therapy selection and follow-up strategies, going beyond what traditional Gleason grading captures.
Radiation therapy: AI-driven contouring algorithms automate the delineation of the prostate and organs at risk (bladder, rectum, femoral heads) for both external beam radiation and brachytherapy. Multicenter studies show these reduce planning time while maintaining high conformity indices, producing reproducible and personalized treatment plans that minimize healthy tissue exposure.
Focal therapy guidance: AI models integrate MRI, clinical, and genomic data to guide minimally invasive focal therapies including high-intensity focused ultrasound (HIFU) and cryotherapy. The AI improves candidate selection by identifying which patients' tumors are well-suited for focal ablation, and enhances targeting accuracy during the procedure itself.
Surgical margin estimation: For robotic-assisted radical prostatectomy, deep learning algorithms analyze multiparametric MRI and histopathology to estimate tumor extent and predict optimal surgical margins. This directly supports the decision of whether nerve-sparing surgery is feasible (preserving urinary and sexual function) or whether wider excision is necessary to ensure complete cancer removal, reducing positive margin rates.
Personalized systemic therapy: For advanced and metastatic prostate cancer, AI predictors use multi-omics (genomic, transcriptomic, proteomic) and clinical datasets to recommend optimal drug combinations and sequence adjustments. Adaptive machine learning models are being developed for real-time therapy adjustment based on treatment response. AI-assisted platforms also monitor post-treatment imaging and biomarkers for rapid detection of suboptimal responses or complications, triggering earlier clinical intervention.
Validation gaps: Most AI studies in prostate cancer are retrospective and single-center, meaning models were trained and tested on data from the same institution. Performance often degrades significantly when deployed at different hospitals with different patient demographics, MRI vendors, scanning protocols, and pathology workflows. The authors emphasize that device dependence and spectrum bias remain largely unaddressed.
Methodological concerns: Systematic reviews cited in this paper reveal that many studies suffer from methodological flaws and evaluation biases. There is a lack of uniformity across training datasets, algorithms employed, and evaluation metrics used. This makes it difficult to compare results across studies or draw reliable conclusions about which approaches actually work best in clinical practice.
Dataset and equity issues: AI performance is heavily influenced by the specific imaging systems used for data acquisition. Models trained predominantly on data from one scanner vendor or one ethnic population may not generalize. If training sets lack diversity, the AI risks worsening existing health disparities, particularly for populations already underserved by the healthcare system.
Automation bias: There is genuine concern about over-reliance on AI outputs. When clinicians trust the algorithm too readily, diagnostic skill atrophy can occur, and errors become harder to catch. The review notes that clinical integration requires explicit strategies to mitigate automation bias, with different performance thresholds depending on the intended use: high sensitivity/NPV for rule-out triage, and high specificity/PPV for confirmatory decisions.
The authors lay out a clear roadmap for translating these research findings into trustworthy clinical tools. The top priority is prospective, multi-site validation with harmonized datasets and standardized reporting, moving beyond the retrospective single-center designs that dominate the current literature.
They call for interoperable deployment within existing digital pathology and imaging ecosystems, proper governance frameworks for safety and equity, and post-market monitoring after clinical deployment. AI tools should be calibrated for their intended use case, whether that is pre-screen triage (where sensitivity matters most) or confirmatory diagnosis (where specificity is paramount).
Future research priorities include open benchmarks that allow fair comparison across models, integration of multi-omics and longitudinal data to build more comprehensive patient profiles, and clinical trials powered for patient-important outcomes (survival, quality of life) and cost-effectiveness rather than just technical accuracy metrics. The overarching vision is human-in-the-loop AI: systems that augment clinical decision-making rather than replace it, ultimately improving both survival and quality of life for prostate cancer patients.