Artificial intelligence and machine learning are transforming medicine by enabling computers to recognize complex patterns in large datasets -- tasks that would be impractical or inconsistent if performed manually. In prostate cancer care, AI is being applied across every stage, from initial diagnosis to treatment planning and long-term prognosis.
This review, from the University of Southern California, synthesizes the most recent literature on AI in prostate cancer management. The authors systematically searched PubMed-Medline for studies published within the prior four years, focusing on the four areas where research has concentrated: radiomics, pathomics, surgical skill assessment, and patient outcome prediction.
Radiomics refers to the extraction of quantitative features from medical images such as MRI, CT, ultrasound, or PET scans. Pathomics refers to AI analysis of tissue specimens obtained from biopsies or surgery. Both approaches generate objective data that can supplement or eventually replace subjective human interpretation in some contexts.
The overarching message of the review is that AI has demonstrated improved accuracy and efficiency across many prostate cancer tasks, but further research -- particularly prospective clinical validation -- is needed before these tools can be safely integrated into routine clinical practice.
Multiparametric MRI (mpMRI) has become central to prostate cancer diagnosis, and AI-based radiomics can extract quantitative features from these images -- including tumor size, shape, texture, and signal intensity -- to make objective predictions about cancer characteristics. One study trained machine learning classifiers on MRI features and achieved sensitivity of 0.93 for peripheral zone tumors in predicting Gleason grade 4 component, outperforming the average sensitivity of 0.72 achieved by three experienced radiologists.
A major clinical challenge is the ambiguous PI-RADS category 3 lesion designation, which represents equivocal findings on MRI where the likelihood of clinically significant cancer is uncertain. Multiple AI models have shown promise in resolving this ambiguity: one random forest classifier achieved an AUC of 0.76 for predicting clinically significant prostate cancer in PI-RADS 3 lesions, while another LASSO-based model achieved AUC of 0.82 with 88% sensitivity in the training set.
For risk stratification, AI models using PSMA PET-CT data achieved AUCs of 0.86 for predicting lymph node invasion and 0.81 for predicting Gleason score greater than 8 -- performance that matches or exceeds traditional clinical tools for these predictions. Combined PET-MRI approaches have achieved overall patient risk model AUCs as high as 0.94.
Ultrasound-based AI is an emerging area with significant potential. One model combining ultrasound images with clinical data achieved an AUC of 0.835 for detecting high-grade cancer (Gleason grade group 4 or above), while a random forest-based classifier using multiparametric ultrasound achieved an AUC of 0.90 for Gleason greater than 3+4 cancers -- potentially offering MRI-equivalent risk prediction at lower cost.
Pathomics applies AI to the analysis of microscopic tissue images from prostate biopsies or surgical specimens. One of its primary applications is automated Gleason grading -- reducing the well-documented inter-observer variability among pathologists in assigning Gleason scores and grade groups.
Multiple AI systems have demonstrated performance comparable to expert pathologists in Gleason grading tasks. One deep learning system achieved 91.5% accuracy in classifying slides as benign or malignant, with 85.4% accuracy in finer-grained classification across Gleason 3, 4, and 5 patterns. The most challenging distinction was between adjacent Gleason grades (3 vs. 4, and 4 vs. 5), particularly when Gleason 4 presented as small or fused glands without lumina.
A Lancet Oncology population-based diagnostic study demonstrated that AI systems can reliably detect and grade prostate cancer in needle core biopsies at a level comparable to expert pathologists. These systems could screen out benign biopsies and automate the measurement of cancer length in malignant ones -- potentially reducing diagnostic workload and errors simultaneously.
An innovative application converts traditional 2D histopathology slides into 3D computational models using AI, enabling more detailed analysis of how cancer glands are organized in three dimensions. One study using this approach achieved AUCs of 0.99 for cancer detection, 0.91 for grade group classification, and 0.80 for recurrence prediction -- suggesting that 3D spatial analysis captures biologically relevant information invisible in 2D slides.
One of the most distinctive applications of AI in prostate cancer is using surgical robot data and video to objectively evaluate surgeon skill during robot-assisted radical prostatectomy (RARP). Surgical robots generate precise kinematic data -- tracking the path, velocity, and movement patterns of surgical instruments -- that AI models can use to distinguish expert surgeons from novices and predict patient outcomes.
A random forest model using kinematic data achieved 87.2% accuracy in predicting length of hospital stay after RARP. A deep learning model trained on the same data predicted urinary continence recovery with surgeons showing better kinematic performance demonstrating significantly higher continence rates at 3 months (47.5% vs. 36.7%) and 6 months (68.3% vs. 59.2%) post-operatively. Importantly, kinematic performance was found to be a better predictor of outcomes than surgeon experience alone -- challenging the traditional assumption that experience is the best proxy for surgical quality.
AI computer vision models applied to surgical video footage have extended this capability further. A unified deep learning framework achieved AUCs between 0.86 and 0.90 for surgical phase recognition, between 0.68 and 0.90 for gesture classification, and between 0.80 and 0.88 for surgical skill assessment across multiple surgical steps and different procedures. Surgical gestures can now be automatically segmented and categorized without human annotation.
Most remarkably, analysis of surgical gesture sequences during nerve-sparing RARP was used to predict patients' erectile function recovery at one year with AUCs of 0.77 and 0.68 at two international centers -- outperforming prediction based on traditional clinical features alone. This finding opens the possibility of providing patients with personalized predictions of functional outcomes based on intraoperative surgical quality.
AI has been applied to non-surgical prostate cancer treatment as well. A machine learning model trained to generate external beam radiation therapy (RT) plans produced plans that were clinically acceptable in 89% of cases when evaluated by a blinded clinician, and 72% were preferred over human-generated plans in head-to-head comparisons. The median total planning time decreased by 60% (from 118 to 47 hours).
However, the radiation planning study revealed an important clinical reality: when the AI-generated plans moved from simulation to actual deployment where real patient care was at stake, physician acceptance of the superior AI plans dropped by 21%. This underscores that clinical acceptance is not guaranteed by technical performance -- human trust, workflow integration, and clinical culture are independent barriers to AI adoption.
For brachytherapy -- where radioactive seeds are implanted in and around the prostate -- an AI-based segmentation algorithm was developed to automatically delineate the prostate boundary from transrectal ultrasound images. This task is normally done manually and is prone to observer variability. The automated system provides near-instantaneous results, potentially accelerating the planning workflow and improving consistency.
AI has also been used to support shared decision-making between patients and clinicians. A web-based tool called askMUSIC trained a random forest model on data from 7,543 prostate cancer patients to generate personalized treatment recommendations across active surveillance, radical prostatectomy, radiation therapy, and hormone therapy. Such tools empower patients to understand how men with similar profiles have been treated and what outcomes they experienced.
AI's ability to integrate many clinical variables simultaneously makes it a powerful tool for prognosis. A machine learning model trained on 30 clinical features predicted prostate cancer mortality within 10 years of diagnosis with an accuracy of 0.98. The most influential predictors were Gleason score, PSA at diagnosis, and age -- consistent with established clinical knowledge, which adds credibility to the model's predictions.
An online decision-support tool based on a long short-term memory neural network -- a type of AI well-suited for time-series data -- predicted individualized 5-year and 10-year survival outcomes for 7,267 prostate cancer patients based on their initial treatment modality. This level of personalization exceeds what is possible with traditional population-level survival statistics.
AI has also been applied to recurrence prediction after radical prostatectomy. Three machine learning models outperformed traditional clinical nomograms in predicting biochemical recurrence at 1, 3, and 5 years post-surgery, with the best model achieving an AUC of 0.894 for 5-year recurrence prediction. Accurate recurrence prediction is clinically important because it guides decisions about adjuvant therapy timing.
An important equity dimension has emerged from AI prognostic research. A random forest model analyzing the SEER database found that while tumor characteristics at diagnosis are the most important predictor of prostate cancer mortality, race and healthcare access factors had equally significant impacts as race itself. A separate AI model built specifically with an African American stromal gene expression signature outperformed standard nomograms in this population, highlighting both the potential and the responsibility of AI to address health disparities.
Despite impressive performance in research settings, AI models for prostate cancer face significant barriers to clinical implementation. Most critically, AI systems are only as good as the data they are trained on -- models trained on data from one institution or patient population may not perform as well when applied to different patient demographics, imaging protocols, or clinical contexts.
Algorithmic bias is a real and documented concern. AI models trained predominantly on data from one racial or socioeconomic group may systematically underperform for underrepresented groups. Multiple studies have highlighted that race and social determinants of health are strong predictors of prostate cancer outcomes, and AI systems that do not account for these factors may perpetuate existing health inequities.
In pathomics specifically, training models on annotations from a single expert can introduce systematic bias reflecting that expert's individual tendencies. Studies have shown this can be partially addressed by training on annotations from multiple experts and using patient-level rather than patch-level cross-validation -- but this requires assembling larger and more diverse annotated datasets.
The gap between simulated AI performance and real-world clinical acceptance is also important. The radiation planning study found that clinicians accepted fewer AI-generated plans than would have been expected from blinded evaluation, highlighting that technical performance alone does not determine whether AI tools are actually used. Trust, transparency, and workflow integration are as important as accuracy.
AI has demonstrated potential to improve nearly every aspect of prostate cancer management -- from more accurate MRI interpretation and Gleason grading, to faster radiation planning, to objective surgical skill assessment and personalized outcome prediction. The breadth of applications is broader than in almost any other cancer type.
However, the authors are clear that AI currently functions as a decision-support tool for urologists and specialists -- not a replacement for clinical judgment. No AI system has yet demonstrated readiness for fully autonomous clinical decision-making in prostate cancer care, and the ethical implications of algorithmic errors in life-or-death situations require careful attention.
The field particularly needs prospective validation studies -- testing AI models on patients in real clinical workflows, rather than retrospective performance on historical datasets. Prospective validation is the standard required for any diagnostic or prognostic tool before clinical adoption, and few AI systems in prostate cancer have yet met this bar.
As AI tools mature and evidence accumulates, integration into clinical practice will likely be gradual and domain-specific -- first in settings where the benefit is largest and the risk of errors lowest, such as pre-screening MRI reads or surgical training feedback. The goal is augmenting clinician capability, not replacing human judgment, in the management of one of the most prevalent cancers affecting men worldwide.