Prostate cancer is the most commonly diagnosed non-skin cancer in men and the second leading cause of cancer mortality in men. Management requires integrating data from multiple sources including PSA levels, MRI-guided biopsies, genomic biomarkers, and Gleason grading of biopsy tissue -- a complex decision-making process that involves significant subjectivity at nearly every step.
A shortage of urological pathologists relative to the volume of prostate biopsies performed creates a bottleneck in diagnosis, while variability in how pathologists interpret histological slides leads to inconsistent grading of the same tissue. This variability in grading can result in overtreatment of indolent cancers or undertreatment of aggressive ones, both of which harm patients.
Artificial intelligence (AI), specifically artificial neural networks (ANNs), offers a path to standardize and automate these high-volume, high-variability tasks. ANNs learn statistical relationships between inputs and outputs from training data, processing nonlinear patterns across many variables simultaneously -- a capability humans cannot match at scale.
The economic dimension adds further urgency. The 10-year management costs for low-risk and high-risk localized prostate cancer patients have been estimated at $45,957 and $188,928 respectively. AI tools that reduce unnecessary biopsies, improve risk stratification, and support better treatment selection have the potential to reduce this burden substantially while improving outcomes.
This systematic review was conducted according to PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. Three databases were searched: PubMed, Embase, and Web of Science, using a comprehensive search string combining terms for prostate cancer, machine learning, neural networks, and clinical trials or active surveillance.
An initial search returned 341 manuscripts. After removing duplicates and applying an exclusion criterion (removing case reports, opinion pieces, conference papers, animal studies, reviews, and commentaries), 81 full-text articles were assessed for eligibility. The final included set comprised 19 manuscripts covering AI applications across PSA-based detection, histopathologic diagnosis, MRI analysis, biomarker validation, patient-centered tools, and risk stratification.
The review focused specifically on studies using artificial neural network-based approaches, distinguishing them from simpler rule-based or linear statistical models. This scope reflects the growing clinical relevance of deep and shallow neural network architectures as practical tools deployable in urology workflows.
The 19 included studies represent a range of AI types and clinical applications: some used simple multilayer perceptrons for PSA interpretation, while others used more complex convolutional neural networks for image analysis. All were evaluated for their contribution to reducing subjectivity and improving efficiency in prostate cancer diagnosis, risk stratification, or patient management.
Prostate-specific antigen (PSA) remains the most commonly used screening tool for prostate cancer, but its specificity is poor. A PSA above 4 ng/mL is the standard threshold for biopsy referral, yet up to 20% of prostate cancers are never detected using this threshold, and a large proportion of elevated PSA values are caused by benign conditions such as prostatitis or benign prostatic hyperplasia.
Multiple studies have demonstrated that incorporating PSA into neural network models alongside additional variables substantially improves diagnostic accuracy. Djavan et al trained an ANN on 1,246 men with PSA levels in the 2.5 to 10 ng/mL range and showed that at 95% sensitivity, the ANN produced better specificity than any single conventional PSA parameter -- meaning fewer unnecessary biopsies without missing cancers.
Finne et al used the proportion of free PSA (%fPSA) as a key input to an ANN trained on 656 men. At 95% sensitivity, the ANN eliminated 33% of unnecessary biopsies in the borderline PSA group (4 to 10 ng/mL), compared to only 19% with the free PSA fraction alone and 24% with logistic regression. This confirmed that the ANN extracted additional predictive signal from combining variables that no single marker captured independently.
Stephan et al further demonstrated that an ANN incorporating %fPSA, patient age, prostate volume, and digital rectal examination (DRE) findings enhanced the specificity of free PSA by 20 to 22% in men with total PSA between 2 and 10 ng/mL. Consistently across these studies, neural networks outperformed conventional PSA parameters and in most cases also outperformed logistic regression models at equal sensitivity levels.
Histopathological grading using the Gleason scoring system is the gold standard for assessing prostate cancer aggressiveness, but grading variability between pathologists is a well-recognized clinical problem. An early AI study by Bhele et al used a machine learning tool trained on 105 images from 38 radical prostatectomy specimens, achieving 67 to 81% concordance with expert outlines of Gleason grade 3 and grade 4 patterns -- demonstrating the feasibility of automated Gleason grading with limited data.
The most striking histopathology result in the review comes from Strom et al, who trained ANNs on digitized slides from 1,247 men. The system achieved an area under the ROC curve of 0.997 for distinguishing benign from malignant biopsy cores on an independent test set and 0.986 on an external validation dataset. These results were described as comparable to the performance of an international expert in prostate pathology -- a rare benchmark for any automated system.
Waliszewski et al developed an ANN to stratify prostate adenocarcinomas based on the structural complexity of tumor cell nuclei distribution. The system reduced intraobserver variability -- the tendency for the same pathologist to grade the same case differently on repeat review -- which the authors suggested could improve the selection of patients for active surveillance by providing more consistent grade classifications.
Lawrentschuk et al trained both a polychotomous logistic regression (PR) model and an ANN to predict the outcomes of 3,025 biopsies for men with PSA below 10 ng/dL. In this study, the ANN did not outperform logistic regression, and the authors cautioned that neural networks require careful training design and may not universally improve on simpler models. This finding serves as an important counterpoint to the generally positive results in the literature.
Multi-parametric MRI (mpMRI) combining T2-weighted imaging, diffusion-weighted imaging (DWI), and spectroscopy has improved prostate cancer detection and localization, but determining tumor aggressiveness from MRI requires significant expertise and remains subjective. Several studies reviewed here applied machine learning to automate Gleason grade prediction from MRI features.
Fehr et al combined apparent diffusion coefficient (ADC) maps and T2-weighted texture features to classify cancer aggressiveness automatically. The system distinguished Gleason score 6 from higher grades, and differentiated 7(3+4) from 7(4+3) tumors with 93% accuracy in both peripheral zone (PZ) and transition zone (TZ), outperforming ADC measurements alone. Antonelli et al similarly showed that ML classifiers built from quantitative MRI features predicted Gleason pattern 4 presence with greater accuracy than three board-certified radiologists.
Toivonen et al used texture analysis of T2-weighted images, DWI, and T2 relaxation maps from 100 patients to create a classifier distinguishing low-risk Gleason 3+3 cancers from higher grades. Their results confirmed that multi-sequence MRI texture features, particularly those derived from diffusion-weighted images processed with monoexponential and kurtosis models, carry substantial predictive information about tumor aggressiveness that current PI-RADS scoring does not fully capture.
In biomarker validation, Green et al used an ANN to compare the prognostic value of Ki67 and DLX2 gene expression. Both markers independently predicted metastasis formation, but only 6.8% of prostate cancer patients have high Ki67 expression, limiting their utility to a small subgroup. Kim et al used targeted proteomics combined with machine learning to discover new non-invasive blood-based signatures for extracapsular prostate cancer, demonstrating that computationally guided proteomics can uncover clinically useful biomarkers from complex multi-protein datasets.
Current prostate cancer risk stratification per NCCN guidelines is based on TNM stage, Gleason grade, PSA level, and biopsy results, placing patients into seven risk categories that guide treatment decisions ranging from active surveillance to radical prostatectomy. While comprehensive, these guidelines do not account for the risk of post-treatment recurrence, which is one of the most clinically important outcomes for patients.
Kumar et al addressed recurrence prediction by developing two separate convolutional neural networks (CNNs) to analyze hematoxylin and eosin (H&E) stained biopsy images. The first network detected individual cell nuclei; the second classified the tissue patches around each nucleus. The combined output achieved an AUC of 0.81 for recurrence prediction, validated on 30 recurrent and 30 non-recurrent controls -- suggesting deep learning can identify histological features predictive of recurrence that are invisible to conventional grading.
For patient-centered care, Auffenberg et al developed askMUSIC, a web-based registry using a random forest machine learning model trained on data from 45 urology practices. Patients can access the platform and compare their predicted treatment options to those actually chosen by other patients with similar clinical characteristics, reducing anxiety about treatment decisions by grounding them in real-world outcomes data from thousands of comparable cases.
The PRODIGE project takes a different approach to standardization, developing an Umbrella Protocol that formalizes how clinical variables are defined and shared across institutions. This structured ontology creates datasets consistent enough to train AI decision support systems that work across multiple hospitals and practice settings -- addressing one of the core barriers to multi-institutional AI deployment in prostate cancer care.