Abstract: Artificial intelligence (AI) is rapidly changing the field of medicine, and prostate cancer is no exception. The significant heterogeneity that characterizes the natural history of prostate cancer often leads to under- or overtreatment. Moreover, the already substantial burden of prostate cancer care on the health care system is predicted to rise significantly. By discerning patterns within immense, complex datasets, AI has the potential to augment the diagnosis, risk stratification, and treatment of prostate cancer beyond what is possible with existing clinical tools. In recent years, AI algorithms have achieved impressive diagnostic accuracy in the realms of imaging and histopathology interpretation, and AI-based biomarkers for risk stratification have been incorporated into major clinical guidelines. Early strides have also been made in radiation treatment planning, intraoperative surgical assistance and surgical education, and quality control. Moving forward, the prospective validation of novel AI algorithms across large, multi-institutional datasets is needed to minimize bias and ensure validity. Furthermore, it is the responsibility of providers across the continuum of prostate cancer care to ensure the safe and ethical integration of AI into clinical practice. This review summarizes the current state of AI applications in the diagnosis, risk stratification, and treatment of prostate cancer, highlighting recent advances and emerging opportunities in this ever-changing field.
Introduction
Artificial intelligence (AI) is rapidly transforming the landscape of prostate cancer management. Prostate cancer is the most commonly diagnosed cancer and the second highest cause of cancer death among American men,1 and the burden of diagnosis and treatment within the US health care system is only projected to increase.2 Prostate cancer presents a unique set of clinical challenges, including a highly variable disease course, rapidly evolving screening and diagnostic paradigms, and multiple treatment modalities that necessitate personalized and shared decision making. Traditional clinical tools and technologies have proved inadequate in addressing those challenges, leaving ample opportunity for AI-based technologies to improve the care of patients with prostate cancer (Figure).
The technical nuances of AI algorithm architecture and development are beyond the scope of this article. Briefly, machine learning (ML) is a foundational AI approach in which algorithms are developed that extrapolate patterns from data without being explicitly programmed. This process overcomes some of the limitations of predictive models and nomograms based on traditional statistical methods such as multivariable regression, which are limited to clinical variables that can be identified by human individuals. In deep learning (DL) methods, multiple layers of learning are used to extract progressively higher-level data features. DL methods are particularly adept at extracting hierarchical features from complex data such as pathology slides, medical images, and even surgical videos. These AI-based technologies have demonstrated promising capabilities across the spectrum of prostate cancer care, including imaging interpretation, automated pathology grading, surgical quality control, intraoperative assistance, and radiation treatment planning.
The clinical integration of AI holds tremendous potential to meet the needs of the modern patient by leveraging the wealth of data that already guides prostate cancer care. However, many barriers remain before AI can be more widely and effectively adopted. Most prominently, concerns exist about bias and reproducibility stemming from the retrospective nature of most AI studies, data heterogeneity, and the “black box” methodology of many AI algorithms. Regulatory, social, and ethical factors are also important considerations. Herein, we aim to capture the current state of AI technologies across the spectrum of prostate cancer care. This review highlights emerging technologies, clinical advances, and new challenges that will affect all providers who treat prostate cancer in the era of AI.
Initial Diagnosis and Evaluation
Imaging Interpretation
Magnetic resonance imaging (MRI) plays an important role in the initial workup of prostate cancer and is supported by both American and European guidelines.3,4 As the volume of prostate MRIs has skyrocketed in recent years,5 the use of AI for automated MRI interpretation has become a prominent focus of research. The recent Prostate Imaging-Cancer Artificial Intelligence (PI-CAI) study developed an AI system to detect Gleason grade group 2 or higher cancers on MRI.6 This AI system demonstrated both noninferiority and superiority in comparison with 62 human radiologists in isolation when tested on 400 MR images (area under the curve [AUC], 91% vs 86%). When applied to 1000 real-world MRI interpretations by radiologists with access to clinical context and peer discussion, the AI system demonstrated comparable performance, although missing the statistical cutoff for noninferiority. Although prior AI models have demonstrated diagnostic accuracy comparable with that of radiologists,7 the PI-CAI study has one of the largest sample sizes to date.
One limitation of the PI-CAI system is that initial model training still relies on manually segmenting regions of interest. This feature may limit the number of training cases that can be incorporated, a problem that was mitigated by international crowdsourcing in the case of PI-CAI. Other groups have used alternative techniques that do not require manual lesion identification. Cai and colleagues used 1514 MRI examinations to train a convolutional neural network to predict grade group 2 or higher prostate cancers.8 This study showed an AUC of 0.89 in the internal test set and 0.86 upon external validation, similar to the performance of radiologists. Furthermore, Dimitriadis and colleagues incorporated 6458 MRI examinations to develop a classifier predicting the presence of prostate cancer, with an AUC of 0.91.9 Of note, their input consisted of biparametric MRI, whereas many other models rely on multiparametric examinations.
Recently, exploratory studies have suggested that AI could one day exceed the diagnostic performance of radiologists. Roest and colleagues developed an ML classifier that detects prostate cancer by analyzing DL-generated heat maps from biparametric MRI, incorporating both prior and current imaging examinations.10 The diagnostic performance of their longitudinal AI model was superior to that of radiologists (AUC, 81% vs 69%). These initial results support the need for further investigation into AI-powered MRI interpretation algorithms in the prospective setting.
AI applications in prostate cancer imaging extend beyond cancer detection with MRI. For example, MRI segmentation to create 3-dimensional (3D) prostate models is currently done manually, a process that is extremely labor-intensive. Boellaard and colleagues evaluated a commercially available AI algorithm for automatic prostate MRI segmentation, comparing its performance with that of manual segmentation.11 They found that AI-assisted segmentation showed promising accuracy, with only minor adjustments needed in most cases, and reduced the time required from 15 to 30 minutes to 2 to 3 minutes. This finding has significant implications for MRI-based fusion prostate biopsy, as well as the generation of 3D models for surgical planning. In addition, DL systems have been trained to segment the prostate automatically into the peripheral and transition zones. This segmentation facilitates the calculation of zone-specific prostate-specific antigen (PSA) density, which has been shown to improve the prediction of clinically significant prostate cancer as well as the specificity of MRI vs that of traditional PSA density.12,13 Finally, early progress has been made in applying AI to positron emission and computed tomography (PET-CT). When a convolutional neural network was trained to identify sites suspicious for prostate cancer on the basis of prostate-specific membrane antigen (PSMA) ligand PET-CT imaging, the accuracy was approximately 80% for identifying suspicious uptake, localizing anatomic site, and determining cancer stage.14 These novel approaches lay the groundwork for boosting efficiency in imaging interpretation, which will be of invaluable assistance to radiologists handling the ever-increasing volume of prostate cancer imaging studies.
Automated Digital Pathology
The histopathologic assessment of prostate biopsy samples provides another fertile area for AI application. The current pathology process remains labor-intensive and subjective, with large volumes of biopsy specimens placing strain on pathologists and causing notable inter- and intrapathologist variability.15,16 In 2019, Paige Prostate, a multiple-instance learning system, was developed to classify whole-slide images without the need for pixel-level annotation.15 The system made it possible to discriminate accurately between slides with benign and slides with cancer-containing cores, thereby enabling the rapid exclusion of benign slides and reducing pathologist workloads without sacrificing sensitivity.15,17 Paige Prostate received US Food and Drug Administration (FDA) approval in 2021, and several groups have since prospectively evaluated its performance. In a study that examined 600 prostate biopsy cores, the sensitivity and negative predictive value of Paige Prostate approached 100%, and it was able to reduce diagnostic time by 66%.17 Other reported benefits include a reduction in the use of immunohistochemistry, which in turn decreased costs.18
Whereas Paige Prostate is FDA approved to assist pathologists by identifying slides suspicious for malignancy, numerous other AI applications are being explored for their ability to provide more detailed pathologic analysis. The current standard for grading prostate cancer is the Gleason score,19 on which all major clinical guidelines base risk stratification and management strategies.3,4 Intense interest is being shown in using AI to automate the Gleason grading system. In 2020, this subject was addressed in 4 major studies. Ström and colleagues trained a deep neural network that distinguished malignant tissue with an AUC of 99.7% in their internal test dataset and of 98.6% in external validation.16 Furthermore, the algorithm achieved accuracy in determining Gleason score comparable to that of a panel of 23 expert uropathologists. Bulten and colleagues trained a DL system that achieved an AUC of 99% for distinguishing malignant tissue and of 98% for identifying grade group 2 or higher disease.20 Furthermore, the algorithm outperformed 10 of 15 human pathologists. Pantanowitz and colleagues trained a convolutional neural network that achieved an AUC of 99.7% for detecting malignant tissue and maintained an AUC of 99.1% on external validation.21 On external validation, their algorithm had an AUC of 94.1% for identifying cancer with a Gleason score of 7 or higher. Importantly, when they deployed this algorithm within their clinical workflow to trigger an alert for second review when the AI model predicted a high Gleason score but the human pathologist identified benign or low-grade tissue, one case of missed cancer was identified. Nagpal and colleagues trained a DL system specifically to assign a Gleason score.22 When compared with Gleason scores assigned by general pathologists, the scores assigned by their algorithm were significantly closer to scores determined by a panel of expert uropathologists, achieving a concordance of 71% in both internal and external datasets.
Also in 2020, the positive results from these individual studies were further corroborated by the Prostate cANcer graDe Assessment (PANDA) challenge, an international AI pathology competition.23 The organizers asked teams to use a dataset of 10,616 digital biopsy slides to develop algorithms for both diagnosing cancer and assigning a Gleason score. The top-performing algorithms were found to outperform general pathologists diagnostically, and results were comparable with those of experienced uropathologists. When the best algorithms were validated on external biopsy slides, the rate of agreement with expert uropathologists was 86%. The main errors made by the AI algorithms were the overdiagnosis of benign cases and a generalized tendency to overgrade slides.
Finally, AI techniques have shown early efficacy in identifying detailed morphologic features of prostate biopsy cores, including length of cancer involvement,16 perineural invasion,21 and cribriform growth pattern,24 all of which have implications for prognosis and management decisions. In summary, AI-based automated tools for pathology interpretation have achieved nearly perfect sensitivity for distinguishing between benign and malignant tissue. However, the level of concordance between the AI-derived Gleason scores and those of expert uropathologists is not yet high enough for the standalone use of AI. Future trials are needed to assess the performance of AI algorithms prospectively, with a focus on Gleason score assessment. Although AI is unlikely to supplant the work of human pathologists, one can envision AI alleviating pathologists’ workloads, identifying discrepancies and human errors for review, and improving diagnostic accuracy in resource-limited settings without access to expert uropathologists.
Prognostication and Prediction
Risk Stratification and Survival Outcomes
Accurate risk stratification is central to the management of prostate cancer and helps guide clinical decisions regarding active surveillance, definitive therapy, and adjuvant treatment. Traditional risk assessment models based on PSA, Gleason score, T stage, and other clinical features are widely accepted but limited in capturing tumor heterogeneity and guiding individualized treatment.25 To address these limitations, AI has become a powerful tool in enhancing individualized risk stratification and prognostication.
One of the most extensively studied AI tools in this space is the multimodal AI (MMAI) model developed by Esteva and colleagues.26 The developers used a DL framework based on digital histopathology and clinical data collected from 5 randomized phase 3 NRG Oncology trials to predict long-term clinical outcomes for intermediate- to high-risk localized prostate cancer. Their training data included 16,204 pathology slides from 5654 patients. This MMAI model improved discriminatory performance relative to National Comprehensive Cancer Network (NCCN) risk stratifications by 11.4% for metastasis, 13.1% for prostate cancer–specific mortality (PCSM), and 11.5% for overall survival (OS). The model is commercially available as the ArteraAI Prostate Test and is now included as an adjunctive tool within the NCCN guidelines.27 These results were subsequently externally validated in a cohort of 318 patients with high-risk or locally advanced disease from the independent NRG/RTOG 9902 phase 3 trial, with favorable performance in predicting distant metastasis and PCSM.28
Beyond localized disease, the MMAI algorithm has demonstrated prognostic relevance in the metastatic setting. In a study by Markowski and colleagues, the MMAI algorithm was applied to 456 patients with metastatic castration-sensitive prostate cancer enrolled in the CHAARTED trial, effectively predicting OS, clinical progression, and the development of castration resistance.29 Finally, in the largest external validation trial thus far for the MMAI algorithm, Parker and colleagues studied 3167 patients recruited to phase 3 trials in the STAMPEDE platform protocol who had metastatic, node-positive, or locally advanced disease.30 The investigators showed that the MMAI algorithm independently predicted PCSM across all disease stages and treatment groups (androgen deprivation therapy [ADT] ± abiraterone or docetaxel), performing better than any clinical variable alone. By refining prostate cancer risk stratification, the MMAI algorithm could be used to select patients for treatment escalation or de-escalation according to their predicted outcomes.
Surgical Applications
The MMAI algorithm has also been applied to surgical cohorts. One ongoing challenge in the management of localized prostate cancer is patient heterogeneity, even within accepted risk stratification guidelines, as standard clinicopathologic variables remain limited in predicting tumor behavior. In a prospective, real-world radical prostatectomy cohort spanning low-, intermediate-, and high-risk localized disease, the MMAI algorithm was found to be independently associated with biochemical recurrence and adverse surgical pathology (grade group 3 or higher, pathologic stage T3b or higher, node positivity).31 The MMAI algorithm demonstrated predictive value even after controlling for NCCN risk group, although the MMAI and NCCN risk stratifications were highly concordant. Li and colleagues applied the MMAI algorithm (developed in prostate biopsy pathology slides) to radical prostatectomy pathology slides from 1021 patients enrolled in the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial.32 The MMAI algorithm was able to predict both PCSM and OS independently of traditional prognostic variables. Although further prospective validation is needed, these findings hold promise for the preoperative selection of patients for active surveillance or deferred treatment, along with postoperative selection for adjuvant therapy or intensified PSA screening. However, it is noteworthy that the Artera MMAI classifier has been evaluated mainly in a post hoc fashion, and future prospective study remains critical.
Precision Medicine: Predicting Therapeutic Response
Beyond prognostication, AI-based tools are being increasingly studied for their ability to predict therapeutic response, marking a shift toward dynamic, biomarker-guided precision oncology. Two studies have sought to predict the need for ADT in patients with localized prostate cancer who received definitive radiation therapy (RT). Using the previously discussed MMAI algorithm as a starting point, Armstrong and colleagues trained a new algorithm to predict which patients might benefit from the addition of long-term ADT to RT.33 In an external validation cohort, patients who were MMAI biomarker–positive displayed a 14% decrease in 15-year metastatic risk with long-term ADT, whereas no benefit was seen in the biomarker-negative group. Complementing this work, Spratt and colleagues developed an independent AI-based model that integrated clinical and pathologic variables to predict benefit from adding short-term ADT to RT.34 Patients identified as model-positive displayed a hazard ratio of 0.34 for 15-year distant metastasis and of 0.28 for PCSM with the addition of ADT, whereas no benefit was seen in the model-negative group. In the near future, AI-based predictive models may be able to personalize care for patients undergoing definitive RT, allowing clinicians to add short-term, long-term, or no ADT selectively to maximize therapeutic benefit while minimizing the adverse effects of hormonal deprivation.
Active investigation is also ongoing in the space of advanced prostate cancer and salvage therapies. Sabbagh and colleagues developed an ML model based on clinical, pathologic, and treatment variables from 2418 patients to predict the risk of distant metastasis in patients undergoing salvage RT after prostatectomy.35 When this model was later externally validated in 475 patients from 2 separate institutions, its discriminative accuracy was higher than that of a previously published nomogram (AUC, 72% vs 60%). This tool could be useful in guiding follow-up surveillance intervals or treatment intensification for patients undergoing salvage RT.
For nonmetastatic castration-resistant prostate cancer, Feng and colleagues applied the aforementioned MMAI algorithm to 420 patients from the SPARTAN trial, which had previously shown that adding apalutamide (Erleada, Janssen) to ADT improves outcomes in this population.36 The MMAI algorithm predicted survival outcomes and also identified patients more likely to benefit from the addition of apalutamide to ADT. MMAI high-risk patients derived significant benefit in regard to metastasis-free, progression-free, and overall survival when treated with apalutamide, whereas MMAI low-risk patients did not. Similarly, Wang and colleagues used the MMAI model in the setting of oligometastatic castration-sensitive disease. Again, in addition to independently predicting OS and time to castration resistance, the MMAI model was prognostic for an improved response to metastasis-directed therapy in a subset of patients.37
Collectively, these studies underscore the versatility of AI-based histopathology in addition to clinicopathologic variables for refining risk stratification, predicting treatment response, selecting clinical trial populations, and guiding treatment decisions across the prostate cancer continuum. These novel applications represent the early stages of precision medicine in prostate cancer, shifting from static risk stratification to dynamic, individual-level predictions of therapeutic benefit. Future work is needed for prospective validation of the additive benefit of AI-based techniques vis-à-vis traditional risk stratification tools, especially in regard to clinically relevant patient outcomes. Furthermore, the ethical implications of using AI for treatment selection mandates careful thought before it is integrated into precision oncology workflows.
Treatment
Surgical Management
Robot-assisted laparoscopic prostatectomy (RALP) is the most widely used surgical treatment for prostate cancer and is a relatively nascent area of AI adoption. In the preoperative setting, ML methods can facilitate the development of complex models to predict postoperative outcomes, incorporating a greater number of clinical features than can be done with traditional statistical techniques. These ML models have been shown to predict postoperative incontinence,38 biochemical recurrence,39 and extraprostatic extension,40 often outperforming logistic regression models developed with the same data. AI can also assist with surgical planning. Mei and colleagues trained a neural network to analyze spatial relationships between the prostate and surrounding pelvic anatomy on the basis of MRI.41 These relationships can then estimate the “difficulty” of RALP as measured by blood loss and operative time.
The vast majority of prostatectomies are performed with an endoscopic camera, generating a wealth of visual data perfectly suited for AI analysis. Indeed, our group has collaborated with Theator, the developer of a surgical intelligence platform, to formulate an AI algorithm that automatically segments RALP operative videos into individual steps with 93% accuracy in comparison with manual human segmentation.42 This algorithm has been applied to annotate live surgeries in real time, including critical views of safety and major intraoperative events.43 We have also used the algorithm to automatically generate RALP operative reports based on the AI-segmented surgical steps.44 These reports were found to be more accurate than surgeon-generated reports (87% vs 73%), with “accurate” defined as free of any clinically significant discrepancy between the report and the surgical video, which was manually reviewed to determine the “ground truth.” We hypothesize that this finding reflects the inaccurate use of prepopulated templates, which affects both billing and clinical care. Further applications of the segmentation algorithm are under active exploration: automating surgical data collection for more accurate billing and research, anticipating operative time and operating room utilization, facilitating surgeon training and quality control, and predicting postoperative outcomes.
In addition to visual data, kinematic data from robotic instruments are a fruitful area of study. Andrew Hung and colleagues have extensively described automated performance metrics (APMs) based on the movement and actions of the robotic arms as tracked by the DaVinci robotic surgical platform (eg, path length, movement time, idle time, energy usage).45 The APMs have subsequently been combined with clinical features and incorporated into DL models to predict outcomes after prostatectomy. The models have been moderately successful in predicting length of stay and continence (most concordance indices are approximately 0.6-0.7).45-47
APMs related to the vesicourethral anastomosis were expectedly most important for predicting continence. Notably, DL models consistently outperformed traditional statistical models with the same inputs. However, in attempts to predict positive surgical margins, the APMs did not add significantly to clinical features alone.48 The authors note that APMs are likely surrogate measures for surgeon expertise and surgical quality, but it remains unclear whether they translate into actionable feedback for performance improvement. This group recently described surgical “gestures” as a purely video-based metric that may offer more clinically relevant classification, showing that the frequency of certain gestures (eg, “hot cut” and “peel/push”) correlate with erectile recovery when used in ML models.49
Intraoperative applications to assist surgeons in real time is perhaps the most exciting AI frontier in RALP. A neural network model based on surgical video predicted intraoperative bleeding events during RALP with 91% accuracy.50 Moreover, initial testing suggests that the model can forewarn surgeons of bleeding events in the coming seconds. Although prospective validation is needed, such warnings could prove to be an invaluable tool for intraoperative patient safety. Porpiglia and colleagues developed a system to superimpose a 3D augmented reality prostate model on the endoscopic video feed during RALP, with the model responding in real time to tissue deformation. Impressively, the technique allowed surgeons to identify microscopic capsular involvement accurately in 15 of 15 cases, whereas only 8 of 17 cases were identified without the augmented reality model.51 However, their initial technique required constant intraoperative manipulation of the 3D model by an additional assistant, thereby limiting practical application in real-world practice. Subsequently, the authors were able to use a convolutional neural network to automate this process. They demonstrated that the AI-generated augmented reality overlay identified the location of neurovascular bundle invasion accurately in 88% of patients with extraprostatic disease,52 potentially allowing precisely targeted nerve sparing in those with high-risk tumors. Finally, the recent NeuroSAFE study demonstrated the utility of intraoperative frozen section analysis of the neurovascular bundle, albeit at the expense of a 50% increase in operative time.53 Could this be mitigated by harnessing AI-driven digital pathology and augmented reality overlay to identify the area of the neurovascular bundle at highest risk more efficiently? These AI techniques may herald a new era of ultraprecise RALP for enhanced nerve sparing even in patients with advanced disease, which could significantly improve sexual recovery and quality of life while still maintaining durable oncologic control.
Radiation Oncology
The feasibility of AI adoption in prostate cancer radiotherapy has steadily increased, with emerging studies demonstrating use in automated segmentation, dose adjustment, and treatment planning. Image segmentation and autocontouring are crucial initial steps for radiation planning. In a prospective clinical evaluation by Tillman-Schwartz and colleagues, a DL model accurately contoured the prostate and surrounding organs. Although human evaluators could still distinguish between AI- and human-generated contours in most cases, the evaluators found efficiency gains in 98% of cases when AI segmentation was used as a starting point, estimating that this saved 17 minutes per patient.54 Similarly, assessment of a commercial DL model showed that AI-generated contours of the prostate and surrounding organs were “acceptable” in 96% of cases and “perfect” in 35%, meeting clinically adequate dose profiles.55 A broader implementation study by Cha and colleagues supported these findings, showing that AI-based automated segmentation reduced segmentation time by an average of 12 minutes and maintained a high level of geometric efficacy.56 However, major edits were required in 33% of cases, highlighting the need for quality oversight. Dosimetric adjustments are also important for maintaining treatment efficacy while minimizing collateral damage, and the use of AI to automate dose adjustments could enhance the efficiency and reproducibility of this process. For example, Rayn and colleagues demonstrated that an AI-driven planning platform could generate dose-escalated stereotactic body radiotherapy plans in 8 to 12 minutes.57
Finally, McIntosh and colleagues tested ML as a tool for developing a curative intent RT plan in a blinded study that included prospective integration into a clinical workflow.58 Their results showed that AI produced acceptable treatment plans in 89% of cases. In addition, 72% of the AI-generated plans were selected over human-generated plans, and plan development time was reduced by 60%. The fact that the number of AI-generated plans ultimately selected was lower than the number of plans deemed acceptable implied some degree of hesitancy in adopting AI for clinical use—possibly because of a lack of robust clinical validation, skepticism regarding accuracy, or concerns regarding accountability and medicolegal implications. Collectively, these data support the mounting applicability of AI techniques to prostate RT, although further improvement and prospective validation of the models are needed. Although AI-generated autocontours and treatment plans still require human supervision, they can provide valuable assistance to increase the efficiency of clinical workflows and enhance reproducibility.
Limitations of AI in Prostate Cancer
Several methodologic limitations and ethical concerns remain concerning the adoption of AI in the management of prostate cancer. The quality of AI models and their outputs are limited by the quality of the input data. Most AI algorithms in prostate cancer were developed by using data from large, tertiary referral centers in Western nations with predominantly White patient populations. Consequently, generalizability to other populations may be limited,59 especially to African American patients, who have historically been underrepresented in prostate cancer research. In addition to compromising external validity, this underrepresentation raises concerns about exacerbating longstanding racial and socioeconomic disparities in prostate cancer care.
Specifically regarding imaging techniques, there is known variability in MRI quality across different MRI scanners, magnet field strength, and radiology imaging protocols. The performance of AI models trained under narrow technical conditions, as is likely in single-institution datasets, may be compromised when the models are applied to outside populations.60,61
Pathologic examination is relied upon as the “ground truth” for most clinical and research applications. Despite efforts to use the Gleason score to standardize pathologic grading, interobserver variability remains substantial.62 Eventually, AI algorithms do hold promise to mitigate this variability, but at present, inconsistencies in the human determination of “ground truth” can lead to inaccuracies in AI-enabled automated pathology, as well as imaging interpretation that requires pathologic confirmation of cancer detection.
The true extent of the above limitations is unknown, and prospective validation is the ultimate test of AI performance. Nearly all AI models are evaluated by internal cross-validation or retrospective validation with an external dataset, which can lead to overfitting.63 It is imperative to incorporate promising AI models into future randomized prostate cancer trials, and to design pragmatic trials specifically to evaluate the real-world effect of AI adoption on patient outcomes.64
Finally, the widespread incorporation of AI into cancer care generates considerable ethical ambiguity. How much should high-stakes clinical decisions, such as whether to perform a biopsy and whether to give or withhold treatment, rely on AI outputs? Although most imaging and pathology algorithms are designed to be conservative, even a small number of missed cancer diagnoses would be unacceptable. The relative performance of algorithmic AI determination vs heuristic human experience and judgment remains unknown.65 Furthermore, uncertainties abound regarding medicolegal accountability and liability with AI utilization. These unanswered questions mandate the development of robust regulatory oversight and rigorous quality control when AI techniques are integrated into clinical practice.
Conclusion
AI research and applications in prostate cancer have blossomed in recent years. In the diagnosis of prostate cancer, the performance of AI algorithms in interpreting imaging and histopathology findings has been comparable with that of physicians. Far from supplanting the roles of radiologists and pathologists, these technologies hold promise for improving efficiency and accuracy as the prostate cancer volume continues to rise. AI-based tools for risk stratification may be the beginning of a revolution in precision oncology across the myriad complex disease states of prostate cancer. Although still in their early stages, AI applications in surgical and radiation treatment may improve quality control, expedite clinical throughput, and mitigate the side effects of treatment. The future of prostate cancer care depends on high-quality prospective trials of novel AI technologies as well as a multidisciplinary ethical and regulatory framework for clinical integration.
Disclosures
The authors have no relevant disclosures.
References
1. Siegel RL, Kratzer TB, Giaquinto AN, Sung H, Jemal A. Cancer statistics, 2025. CA Cancer J Clin. 2025;75(1):10-45.
2. James ND, Tannock I, N’Dow J, et al. The Lancet Commission on prostate cancer: planning for the surge in cases. Lancet. 2024;403(10437):1683-1722.
3. Wei JT, Barocas D, Carlsson S, et al. Early detection of prostate cancer: AUA/SUO guideline part II: considerations for a prostate biopsy. J Urol. 2023;210(1):54-63.
4. Cornford P, van den Bergh RCN, Briers E, et al. EAU-EANM-ESTRO-ESUR-ISUP-SIOG guidelines on prostate cancer-2024 update. Part I: screening, diagnosis, and local treatment with curative intent. Eur Urol. 2024;86(2):148-163.
5. Soerensen SJC, Li S, Langston ME, Fan RE, Rusu M, Sonn GA. Trends in pre-biopsy MRI usage for prostate cancer detection, 2007-2022. Prostate Cancer Prostatic Dis. 2025;28(2):519-522.
6. Saha A, Bosma JS, Twilt JJ, et al; PI-CAI consortium. Artificial intelligence and radiologists in prostate cancer detection on MRI (PI-CAI): an international, paired, non-inferiority, confirmatory study. Lancet Oncol. 2024;25(7):879-887.
7. Roest C, Fransen SJ, Kwee TC, Yakar D. Comparative performance of deep learning and radiologists for the diagnosis and localization of clinically significant prostate cancer at MRI: a systematic review. Life (Basel). 2022;12(10):1490.
8. Cai JC, Nakai H, Kuanar S, et al. Fully automated deep learning model to detect clinically significant prostate cancer at MRI. Radiology. 2024;312(2):e232635.
9. Dimitriadis A, Kalliatakis G, Osuala R, et al. Assessing cancer presence in prostate MRI using multi-encoder cross-attention networks. J Imaging. 2025;11(4):98.
10. Roest C, Kwee TC, Saha A, Fütterer JJ, Yakar D, Huisman H. AI-assisted biparametric MRI surveillance of prostate cancer: feasibility study. Eur Radiol. 2023;33(1):89-96.
11. Boellaard TN, van Erck R, van der Graaf SH, et al. Comparing AI and manual segmentation of prostate MRI: towards AI-driven 3D-model-guided prostatectomy. Diagnostics (Basel). 2025;15(9):1141.
12. Hamm CA, Baumgärtner GL, Padhani AR, et al. Reduction of false positives using zone-specific prostate-specific antigen density for prostate MRI-based biopsy decision strategies. Eur Radiol. 2024;34(10):6229-6240.
13. Kuanar S, Cai J, Nakai H, et al. Transition-zone PSA-density calculated from MRI deep learning prostate zonal segmentation model for prediction of clinically significant prostate cancer. Abdom Radiol (NY). 2024;49(10):3722-3734.
14. Capobianco N, Sibille L, Chantadisai M, et al. Whole-body uptake classification and prostate cancer staging in 68Ga-PSMA-11 PET/CT using dual-tracer learning. Eur J Nucl Med Mol Imaging. 2022;49(2):517-526.
15. Campanella G, Hanna MG, Geneslaw L, et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat Med. 2019;25(8):1301-1309.
16. Ström P, Kartasalo K, Olsson H, et al. Artificial intelligence for diagnosis and grading of prostate cancer in biopsies: a population-based, diagnostic study. Lancet Oncol. 2020;21(2):222-232.
17. da Silva LM, Pereira EM, Salles PGO, et al. Independent real-world application of a clinical-grade automated prostate cancer detection system. J Pathol. 2021;254(2):147-158.
18. Flach RN, van Dooijeweert C, Nguyen TQ, et al. Prospective clinical implementation of Paige Prostate Detect artificial intelligence assistance in the detection of prostate cancer in prostate biopsies: CONFIDENT P trial implementation of artificial intelligence assistance in prostate cancer detection. JCO Clin Cancer Inform. 2025;9(9):e2400193.
19. Epstein JI, Egevad L, Amin MB, Delahunt B, Srigley JR, Humphrey PA; Grading Committee. The 2014 International Society of Urological Pathology (ISUP) consensus conference on Gleason grading of prostatic carcinoma: definition of grading patterns and proposal for a new grading system. Am J Surg Pathol. 2016;40(2):244-252.
20. Bulten W, Pinckaers H, van Boven H, et al. Automated deep-learning system for Gleason grading of prostate cancer using biopsies: a diagnostic study. Lancet Oncol. 2020;21(2):233-241.
21. Pantanowitz L, Quiroga-Garza GM, Bien L, et al. An artificial intelligence algorithm for prostate cancer diagnosis in whole slide images of core needle biopsies: a blinded clinical validation and deployment study. Lancet Digit Health. 2020;2(8):e407-e416.
22. Nagpal K, Foote D, Tan F, et al. Development and validation of a deep learning algorithm for Gleason grading of prostate cancer from biopsy specimens. JAMA Oncol. 2020;6(9):1372-1380.
23. Bulten W, Kartasalo K, Chen PC, et al; PANDA challenge consortium. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge. Nat Med. 2022;28(1):154-163.
24. Ambrosini P, Hollemans E, Kweldam CF, Leenders GJLHV, Stallinga S, Vos F. Automated detection of cribriform growth patterns in prostate histology images. Sci Rep. 2020;10(1):14904.
25. Daskivich TJ, Wood LN, Skarecky D, Ahlering T, Freedland S. Limitations of the National Comprehensive Cancer Network® (NCCN®) guidelines for prediction of limited life expectancy in men with prostate cancer. J Urol. 2017;197(2):356-362.
26. Esteva A, Feng J, van der Wal D, et al; NRG Prostate Cancer AI Consortium. Prostate cancer therapy personalization via multi-modal deep learning on randomized phase III clinical trials. NPJ Digit Med. 2022;5(1):71.
27. Spratt DE, Srinivas S, Schaeffer EM, Adra N, Ahmed B, An Y. NCCN Clinical Practice Guidelines in Oncology: Prostate Cancer. Version 2.2025. Accessed August 11, 2025. https://www.nccn.org/professionals/physician_gls/pdf/prostate.pdf.
28. Ross AE, Zhang J, Huang HC, et al. External validation of a digital pathology-based multimodal artificial intelligence architecture in the NRG/RTOG 9902 phase 3 trial. Eur Urol Oncol. 2024;7(5):1024-1033.
29. Markowski MC, Ren Y, Tierney M, et al. Digital pathology-based artificial intelligence biomarker validation in metastatic prostate cancer. Eur Urol Oncol. 2025;8(3):755-762.
30. Parker CTA, Mendes L, Liu VYT, et al; STAMPEDE collaborators. External validation of a digital pathology-based multimodal artificial intelligence-derived prognostic model in patients with advanced prostate cancer starting long-term androgen deprivation therapy: a post-hoc ancillary biomarker study of four phase 3 randomised controlled trials of the STAMPEDE platform protocol. Lancet Digit Health. 2025;7(7):100885.
31. Bjartell A, Krzyzanowska A, Liu VYT, et al. Validation of a digital pathology-based multimodal artificial intelligence biomarker in a prospective, real-world prostate cancer cohort treated with prostatectomy. Clin Cancer Res. 2025;31(8):1546-1553.
32. Li EV, Ren Y, Griffin J, et al. An artificial intelligence-digital pathology algorithm predicts survival after radical prostatectomy from the Prostate, Lung, Colorectal, and Ovarian Cancer Trial. J Urol. 2025;213(5):600-608.
33. Armstrong AJ, Liu VYT, Selvaraju RR, et al. Development and validation of an artificial intelligence digital pathology biomarker to predict benefit of long-term hormonal therapy and radiotherapy in men with high-risk prostate cancer across multiple phase III trials. J Clin Oncol. 2025;43(32):3494-3504.
34. Spratt DE, Tang S, Sun Y, et al. Artificial intelligence predictive model for hormone therapy use in prostate cancer. NEJM Evid. 2023;2(8):a2300023.
35. Sabbagh A, Tilki D, Feng J, et al. Multi-institutional development and external validation of a machine learning model for the prediction of distant metastasis in patients treated by salvage radiotherapy for biochemical failure after radical prostatectomy. Eur Urol Focus. 2024;10(1):66-74.
36. Feng FY, Smith MR, Saad F, et al. Digital pathology-based multimodal artificial intelligence scores and outcomes in a randomized phase III trial in men with nonmetastatic castration-resistant prostate cancer. JCO Precis Oncol. 2025;9(9):e2400653.
37. Wang JH, Deek MP, Mendes AA, et al. Validation of an artificial intelligence-based prognostic biomarker in patients with oligometastatic castration-sensitive prostate cancer. Radiother Oncol. 2025;202:110618.
38. Amparore D, De Cillis S, Alladio E, et al. Development of machine learning algorithm to predict the risk of incontinence after robot-assisted radical prostatectomy. J Endourol. 2024;38(8):871-878.
39. Ekşi M, Evren İ, Akkaş F, et al. Machine learning algorithms can more efficiently predict biochemical recurrence after robot-assisted radical prostatectomy. Prostate. 2021;81(12):913-920.
40. Kwong JCC, Khondker A, Meng E, et al. Development, multi-institutional external validation, and algorithmic audit of an artificial intelligence-based side-specific extra-prostatic extension risk assessment tool (SEPERA) for patients undergoing radical prostatectomy: a retrospective cohort study. Lancet Digit Health. 2023;5(7):e435-e445.
41. Mei H, Wang Z, Zheng Q, et al. Deep learning for predicting difficulty in radical prostatectomy: a novel evaluation scheme. Urology. 2025;198:1-7.
42. Khanna A, Antolin A, Bar O, et al. Automated identification of key steps in robotic-assisted radical prostatectomy using artificial intelligence. J Urol. 2024;211(4):575-584.
43. Zuluaga L, Rich JM, Gupta R, et al. AI-powered real-time annotations during urologic surgery: the future of training and quality metrics. Urol Oncol. 2024;42(3):57-66.
44. Khanna A, Wolf T, Frank I, et al. Enhancing accuracy of operative reports with automated artificial intelligence analysis of surgical video. J Am Coll Surg. 2025;240(5):739-746.
45. Hung AJ, Chen J, Che Z, et al. Utilizing machine learning and automated performance metrics to evaluate robot-assisted radical prostatectomy performance and predict outcomes. J Endourol. 2018;32(5):438-444.
46. Hung AJ, Chen J, Ghodoussipour S, et al. A deep-learning model using automated performance metrics and clinical features to predict urinary continence recovery after robot-assisted radical prostatectomy. BJU Int. 2019;124(3):487-495.
47. Trinh L, Mingo S, Vanstrum EB, et al. Survival analysis using surgeon skill metrics and patient factors to predict urinary continence recovery after robot-assisted radical prostatectomy. Eur Urol Focus. 2022;8(2):623-630.
48. Lee RS, Ma R, Pham S, et al. Machine learning to delineate surgeon and clinical factors that anticipate positive surgical margins after robot-assisted radical prostatectomy. J Endourol. 2022;36(9):1192-1198.
49. Ma R, Ramaswamy A, Xu J, et al. Surgical gestures as a method to quantify surgical performance and predict patient outcomes. NPJ Digit Med. 2022;5(1):187.
50. Checcucci E, Piazzolla P, Marullo G, et al. Development of bleeding artificial intelligence detector (BLAIR) system for robotic radical prostatectomy. J Clin Med. 2023;12(23):7355.
51. Porpiglia F, Checcucci E, Amparore D, et al. Three-dimensional elastic augmented-reality robot-assisted radical prostatectomy using hyperaccuracy three-dimensional reconstruction technology: a step further in the identification of capsular involvement. Eur Urol. 2019;76(4):505-514.
52. Checcucci E, Piana A, Volpi G, et al. Three-dimensional automatic artificial intelligence driven augmented-reality selective biopsy during nerve-sparing robot-assisted radical prostatectomy: a feasibility and accuracy study. Asian J Urol. 2023;10(4):407-415.
53. van der Slot MA, den Bakker MA, Tan TSC, et al; Anser Prostate Cancer Network. NeuroSAFE in radical prostatectomy increases the rate of nerve-sparing surgery without affecting oncological outcome. BJU Int. 2022;130(5):628-636.
54. Tillman-Schwartz E, Tillman G, Tansky JY, Paly JJ, Efstathiou JA, Wang Y. Prospective analysis of a deep learning auto-contouring model for definitive radiation of localized prostate cancer. Int J Radiat Oncol Biol Phys. 2022;114(3)(suppl):e105-e106.
55. Duan J, Bernard M, Downes L, et al. Evaluating the clinical acceptability of deep learning contours of prostate and organs-at-risk in an automated prostate treatment planning process. Med Phys. 2022;49(4):2570-2581.
56. Cha E, Elguindi S, Onochie I, et al. Clinical implementation of deep learning contour autosegmentation for prostate radiotherapy. Radiother Oncol. 2021;159:1-7.
57. Rayn K, Elliston C, Savacool M, et al. Physician-driven artificial intelligence enabled planning for intraprostatic dose escalation in under ten minutes [ASCO GU abstract 187]. J Clin Oncol. 2023;41(6)(suppl).
58. McIntosh C, Conroy L, Tjong MC, et al. Clinical integration of machine learning for curative-intent radiation treatment of patients with prostate cancer. Nat Med. 2021;27(6):999-1005.
59. Willemink MJ, Koszek WA, Hardell C, et al. Preparing medical imaging data for machine learning. Radiology. 2020;295(1):4-15.
60. Alis D, Onay A, Colak E, Karaarslan E, Bakir B. A narrative review of artificial intelligence in MRI-guided prostate cancer diagnosis: addressing key challenges. Diagnostics (Basel). 2025;15(11):1342.
61. Roberts M, Driggs D, Thorpe M, et al. Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and CT scans. Nat Mach Intell. 2021;3:199-217.
62. Egevad L, Ahmad AS, Algaba F, et al. Standardization of Gleason grading among 337 European pathologists. Histopathology. 2013;62(2):247-256.
63. Park SH, Han K. Methodologic guide for evaluating clinical performance and effect of artificial intelligence technology for medical diagnosis and prediction. Radiology. 2018;286(3):800-809.
64. Kelly CJ, Karthikesalingam A, Suleyman M, Corrado G, King D. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. 2019;17(1):195.
65. Liu X, Faes L, Kale AU, et al. A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: a systematic review and meta-analysis. Lancet Digit Health. 2019;1(6):e271-e297.