Artificial intelligence for active surveillance decision support in patients with prostate cancer: A scoping review

Abstract

Background:

Active surveillance (AS) is an established management strategy for patients with low- and favorable intermediate-risk prostate cancer (PCa). Accurately distinguishing indolent from aggressive disease and identifying progression during follow-up remain significant clinical challenges. Artificial intelligence (AI) has appeared as a promising approach to enhance risk stratification and support clinical decision-making. Our aim was to evaluate the role of AI-based decision support systems in patient selection and monitoring during AS for PCa, as well as their potential to guide timely transition to definitive treatment.

Methods:

A scoping review was conducted in accordance with Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) 2020 guidelines. The review protocol was prospectively registered in PROSPERO (CRD420261358519). Literature searches in PubMed and Scopus were performed, applying filters for English-language studies published within the past 5 years. Of 133 records identified, 24 studies met the inclusion criteria after screening and eligibility assessment.

Results:

Most included studies were retrospective and used machine learning (ML) or deep learning (DL) techniques on clinical, radiological, and histopathological data. AI models demonstrated stronger predictive performance than traditional statistical methods, with reported area under the curve (AUC) values ranging from 0.74 to 0.96. Radiomics-based models derived from magnetic resonance imaging (MRI) facilitated non-invasive characterization of tumor aggressiveness, while DL algorithms applied to biopsy whole-slide images improved grading accuracy and progression prediction. Multimodal models that integrated heterogeneous data sources consistently outperformed single-modality approaches. Finally, longitudinal models including serial prostate-specific antigen (PSA) measurements and imaging data enabled dynamic risk assessment and customized monitoring during AS.

Conclusion:

AI-based models show considerable potential to improve patient selection and predict progression in AS for PCa, potentially reducing overtreatment and unnecessary biopsies. However, current evidence is limited by retrospective study designs, lack of external validation, and variability in AS protocols. Prospective multicenter studies and enhanced model explainability are necessary before routine clinical implementation.

Level of evidence:

Not applicable.

Keywords

prostatic neoplasms [MeSH]prostate cancer prostate carcinoma active surveillance active monitoring artificial intelligence [MeSH]machine learning [MeSH]deep learning radiomics risk stratification predictive modeling

Introduction

Prostate cancer (PCa) is one of the most diagnosed malignancies in men worldwide.¹ Its marked clinical heterogeneity makes management challenging. The disease can behave indolently or aggressively, creating an urgent need for strategies that separate patients who need intervention from those who can be safely observed. Active surveillance (AS) has become a leading approach for patients with low-risk and some favorable intermediate-risk PCa. This strategy seeks to balance oncological safety with minimizing overtreatment.

Although AS is widely used, clinicians must accurately identify which patients can be safely monitored and which may require earlier treatment due to the risk of progression. Current protocols rely on several tools: digital rectal examinations (DREs), prostate-specific antigen (PSA) and PSA density (PSAd) measurements, multiparametric magnetic resonance imaging (mpMRI), and serial prostate biopsies.² These tools may be essential for monitoring, but each has significant limitations. Repeated prostate biopsies are associated with risks, such as bleeding and infections that could require hospitalization. Magnetic resonance imaging (MRI) scans are costly and may not be widely available. In addition, common clinical measures and imaging may not fully capture the aggressiveness of the malignancy, leading to misclassification or delayed detection of cancer progression.

In this context, artificial intelligence (AI) has rapidly emerged as a promising tool in medical research and clinical practice. Unlike conventional methods, AI, through machine learning (ML), deep learning (DL), and radiomics, has the potential to uncover subtle, complex patterns in clinical and imaging data that might be overlooked by human observers.³ This unique analytical capability enables more refined risk stratification and supports clinicians by synthesizing multifaceted data sources to inform decision-making. AI-based models may thus uniquely enhance the identification of suitable candidates for AS and the early detection of disease progression.

This scoping review examines whether AI-based decision support systems can improve patient selection and monitoring during AS for PCa. In addition, it discusses their potential to guide timely moves from surveillance to surgery.

Materials and methods

Two independent reviewers (D.D. and N.A.) performed the literature search through the PubMed and Scopus databases. The review protocol was prospectively registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD420261358519). The search strategy included both Medical Subjects Headings (MeSH) and free-text terms, using the following keywords: “prostatic neoplasms [MeSH],” “prostate cancer,” “prostate carcinoma,” “active surveillance,” “active monitoring,” “artificial intelligence [MeSH],” “machine learning [MeSH],” “deep learning,” “radiomics,” “risk stratification,” and “predictive modeling.” The search was conducted in accordance with Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) 2020 guidelines.⁴ Discrepancies were resolved after consensus with the senior author (T.T.). A total of 133 records were identified. We selected English-written studies published during the last 5 years. Specifically, 63 records were identified in PubMed and 70 in Scopus. Eighteen duplicate records were removed. During the screening process, 89 records were excluded; 53 records were not original articles (reviews, notes, letters, editorials), 20 records were excluded as irrelevant to AI, 13 to AS, and 3 to PCa. A total of 26 reports were assessed for eligibility. After the eligibility assessment and exclusions (two reports were excluded; one was irrelevant to AS and one was published over 5 years ago), 24 studies were included in the review. Figure 1 shows details of the selection process. Due to study heterogeneity and the nonstandardized quality appraisal, a narrative synthesis was performed.

Figure 1.

Literature search flowchart according to PRISMA 2020 guideline.

Results

Twenty-four studies met our inclusion criteria. In most studies, AI models were evaluated on retrospective cohorts of patients with PCa managed on AS.^{5,6,7–12,13,14} Cohort sizes varied from 42¹⁵ to 1422 patients.¹⁶ Predictive models for PCa progression and risk stratification were developed using various ML techniques, most commonly random forests, support vector machines, and decision trees (DTs),^{6,7,11,13,16–22} as well as DL models, including convolutional and recurrent neural networks.^{5,9,19,23–27} Long short-term memory^8,12,14 and temporal outcome-oriented clustering methods¹⁰ were implemented in prospective longitudinal models for predicting progression. Data sources varied across studies and included clinical variables (PSA measurements, PSAd, DRE results, etc.),^{6,7,10,12–14,16,17,19,21,22,25} radiological features extracted from bi- and multiparametric MRI scans,^{7–15,16,18–20,22,24,27,28} and histopathological data.^{5–7,10,17,21,23,25,26} Model performance was mostly reported using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity.

AI for selection of patients eligible for active surveillance

The first major domain, presented in Table 1, included studies that assessed the role of AI in improving risk stratification and identifying patients who could be managed on AS. These studies aimed to distinguish patients with indolent disease from those harboring clinically significant or more aggressive PCa.

Table 1.

Studies that assessed AI in identifying candidates for AS.

Authors	Type of study/patients included	Objective	Variables used for model development	AI methods	Results
Ding et al.⁵	Retrospective cohort (two independent AS cohorts) validation study/138 and 169	To assess whether a DL algorithm could regrade initial prostate biopsy slides and improve risk stratification for AS	WSIs for prostate needle biopsies	DL-based pathology algorithm	Upgrading of the initial biopsy to ⩾ GG2 by AI was associated with rapid or extreme grade reclassification during AS (OR: 3.3, p = 0.04), whereas upgrading by contemporary uropathologist review was not. In the validation cohort, upgrading of the initial biopsy by AI was associated with grade reclassification during AS (HR: 1.7, p = 0.03)
Saitta et al.¹⁶	Retrospective cohort/1422	To develop an ML-based model combining clinical and radiological findings to predict the presence of csPCa	Age, BMI, PCa familiarity, prebiopsy PSA, abnormal DRE, PSAd, prostate volume, type of biopsy, ISUP GG, PRI-MUS score, PI-RADS score, and location of lesions	ML (classification and regression tree model)	Validation performance: AUC: 0.83 Sensitivity: 78% Specificity: 72% NPV: 81%
ElKarami et al.¹⁷	Retrospective cohort/592	To assess ML models on Gleason score upgrading prediction on confirmatory MRI-guided targeted biopsy in AS candidates	BMI, ethnicity, age, DRE, PSAd, number of pre-MRI biopsies, number of positive cores, and results of MRI-TB	ML (SVM, naïve Bayes, RF, bootstrap aggregating, and adaptive boosting)	Best-performing models: Adaptive boosting; AUC: 0.952 (for upgrading defined as Gleason 3 + 4) RF; AUC: 0.947 (for upgrading defined as Gleason 4 + 3)
Nicoletti et al.¹⁸	Retrospective validation study/108	To evaluate ML techniques on distinguishing low- from high- aggressive PCa	Texture features extracted from bpMRI (ADC and T2WI)	ML (SVM, DT, and ensemble learning algorithms)	AUC: 0.81, sensitivity: 77%, specificity: 93% (best-performing model; DT) The DT model distinguished low-aggressive from high-aggressive PCa with high accuracy
Hiremath et al.¹⁹	Retrospective validation study/592	To develop and validate an integrated nomogram (ClaD) combining DL-based imaging predictions, PI-RADS scoring, and clinical variables to identify csPCa on bpMRI	bpMRI, PSA, prostate volume, lesion volume	DL (convolutional neural network)	ClaD (DL, PI-RADS, and clinical); AUC: 0.81 for identifying csPCa. This significantly outperformed the following: - DIN nomogram (clinical and DL only); AUC: 0.74, p = 0.0005, and - PIN nomogram (clinical and PI-RADS only); AUC: 0.76, p < 0.0001
Shahait et al.²⁸	Retrospective cohort/98 (men who underwent RARP)	To identify MRI-derived radiomic features from periprostatic adipose tissue associated with csPCa and evaluate whether these features could help distinguish indolent from aggressive disease	Pre-operative MRI scans (an optimal biomarker method was used to identify features from 7631 intensity- and texture-based properties)	ML (Bayesian optimization)	The six highest-ranking features were associated with the presence of csPCa. AUC: 0.82 Sensitivity: 95% Specificity: 39%
Liu et al.²⁰	Retrospective validation study/956	To develop and validate an MRI-based radiomics model to identify AS candidates non-invasively	MRI (T2WI) scans	ML (XGBoost, LR, RF, adaptive boosting, and DT)	Best-performing model; - XGBoost model, with an average AUC of 0.803, accuracy of 0.693, sensitivity of 66.8%, and specificity of 84.1%
Tsuneki et al.²³	Retrospective DL pathology study/2285 slides (1321 adenocarcinoma and 964 benign)	To develop a DL model that classifies prostate adenocarcinoma into indolent (applicable for AS) and aggressive (necessary for definitive therapy) on core needle biopsy WSIs	H&E-stained histopathological core needle biopsy specimen slides	DL (convolutional neural network)	- AUC: 0.846 for identifying indolent disease - AUC: 0.980 for identifying aggressive disease These suggest that it would be possible not only to detect adenocarcinoma on biopsy WSIs but also to predict patients’ optimal clinical interventions (AS or definitive therapy)
Wu et al.²⁴	Retrospective cohort/110	To develop a multimodal nomogram that predicts risk stratification in PCa patients using DL and radiomics, and to compare it against a traditional clinical nomogram and independent risk factors	mpMRI scans, PI-RADS scoring, functional subsets of peripheral lymphocytes, prostate volume, and age	DL (convolutional neural network)	The multimodal nomogram demonstrated superior performance over the traditional clinical nomogram and various predictors, including PSA AUC: 0.961 (vs. 0.874) Sensitivity: 93% (vs. 77%) Specificity: 87% (vs. 89%)
Sondermann et al.²¹	Retrospective cohort/1929 multi- disciplinary tumor board recommendations	To evaluate whether ML algorithms can reproduce multidisciplinary tumor board recommendations for PCa management, including decisions regarding: PSMA-PET, conventional imaging, AS, and local therapy	Age, PSA levels, prostate volume, DRE results, ISUP grade, number of positive biopsy cores, and comorbidities from patient history	ML (DT, RF, KNN)	Overall predictive accuracies: KNN: 66.6%, RF: 66.3%, and DT: 62.2% Prediction for local therapy was highly accurate; F1 score: 0.99 vs. F1 scores for AS; RF: 0.77, and DT: 0.73 Performance was lower for less frequent recommendations, such as PSMA-PET and AS
Kim et al.²²	Retrospective cohort/241	To determine whether an AI model based on diffusion basis spectrum imaging (DBSI) can predict csPCa before biopsy, and whether it improves risk discrimination when combined with PI-RADS	The DBSI-based model was compared with available biomarkers (PSA, PSAd, and PI-RADS) for risk discrimination of csPCa, defined as Gleason score > 7	ML (RF, SVM, extreme gradient boosting, and multivariable LR)	- The DBSI-model independently predicted csPCa (OR = 2.04, p < 0.01) - DBSI-model and PSAd + PI-RADS model had similar performance; AUC: 0.863 vs. 0.859 (p = 0.89) - Best-performing combination: DBSI-model with PI-RADS; AUC: 0.894 (p < 0.01) For patients with PI-RADS 1–3, the DBSI-model could have reduced biopsies by 27% while missing 2% of csPCa (compared with biopsy-for-all)

ADC = apparent diffusion coefficient, AI = artificial intelligence, ANN = artificial neural network, AS = active surveillance, AUC = area under the curve, bpMRI = biparametric magnetic resonance imaging, CNN = convolutional neural network, csPCa = clinically significant prostate cancer, DL = deep learning, DRE = digital rectal examination, DT = decision tree, GG = grade group, H&E = hematoxylin and eosin, HR = hazard ratio, ISUP = International Society of Urological Pathology, KNN = k-nearest neighbors, LASSO = least absolute shrinkage and selection operator, LDA = linear discrimination analysis, LR = logistic regression, ML = machine learning, mpMRI = multiparametric magnetic resonance imaging, OR = odds ratio, PCa = prostate cancer, PI-RADS = Prostate Imaging Reporting and Data System, PRECISE = Prostate Cancer Radiological Estimation of Change in Sequential Evaluation, PRI-MUS = prostate risk identification using micro-ultrasound, PSA = prostate-specific antigen, PSAd = prostate-specific antigen density, PSMA-PET = prostate-specific membrane antigen positron emission tomography, RARP = robotic-assisted radical prostatectomy, RF = random forest, SVM = support vector machine, WSIs = whole-slide images

Several studies focused on baseline reclassification using clinical variables and biopsy-derived information. In patients undergoing confirmatory MRI-targeted biopsy, ElKarami et al.¹⁷ showed that ML models could accurately predict upgrading, with models achieving AUC values up to 0.952. Similarly, a DL pathology model applied to initial biopsy slides in two independent AS cohorts improved baseline risk stratification by identifying patients initially classified as GG 1 who were more likely to undergo rapid or extreme grade reclassification.⁵ Tsuneki et al.²³ also demonstrated that DL applied to whole-slide images (WSIs) could distinguish indolent from aggressive prostate adenocarcinoma with high discriminative performance (AUC values of 0.864 for indolent and 0.980 for aggressive disease). Regarding clinical decision support systems and recommendations, Sondermann et al. evaluated ML algorithms on reproducing multidisciplinary tumor board recommendations for PCa management (including AS) using clinicopathological variables. However, the model showed moderate performance in reproducing AS recommendations compared with its performance in predictions regarding local therapy or conventional imaging.²¹

Several of the included studies have tested radiomics models derived from prostate MRI scans. Radiomics analysis enables the extraction of multi-dimensional quantitative imaging features that can reflect tumor heterogeneity and biological aggressiveness. Multiple studies have developed ML models using radiomic features derived from biparametric or multiparametric MRI sequences, such as T2-weighted imaging and apparent diffusion coefficient maps.^18,20,22,28 These radiomics-based approaches have demonstrated promising diagnostic performance in distinguishing indolent from aggressive PCa, thereby supporting non-invasive identification of patients suitable for AS. For example, Nicoletti et al.¹⁸ showed that texture-based radiomic features extracted from bpMRI were used to train classifiers that accurately distinguished low-aggressive from high-aggressive disease, achieving 77% sensitivity and 93% specificity with a DT model. Shahait et al.²⁸ explored radiomic features extracted from periprostatic adipose tissue, demonstrating that non-tumoral biomarkers may also help discriminate csPCa. In another study, Liu et al.²⁰ developed a radiomics model using various ML techniques to identify AS candidates, with the best-performing model achieving an average AUC of 0.803, showing potential for non-invasive identification.

Multimodal AI models combining imaging and clinical variables also showed encouraging results and tended to outperform traditional risk stratification tools. Hiremath et al.¹⁹ developed a nomogram that combined DL-based imaging predictions, PI-RADS scoring, and clinical parameters, demonstrating superior performance compared with models based on fewer modalities. Wu et al.²⁴ similarly reported stronger performance (AUC: 0.961) of a multimodal nomogram incorporating mpMRI, peripheral lymphocyte subsets, and clinical variables compared with a traditional clinical nomogram (AUC: 0.874) for risk classification. Saitta et al.¹⁶ integrated micro-ultrasound features with other clinical, pathological, and imaging data to develop an ML-based model for predicting csPCa. Overall, the proposed models showed statistically significant improvements in predictive performance, achieving AUC values ranging from 0.81 to 0.96.^16,19,24 Kim et al.²² showed that combining a diffusion basis spectrum imaging model with the PI-RADS score accurately detected csPCa (AUC: 0.894), with the potential to reduce biopsies by up to 27% in patients with a PI-RADS score of 1–3, while missing 2% of csPCa. These findings underscore the importance of integrating clinical, radiological, and pathological features into predictive models, as this approach achieves higher accuracy and could pave the way for more robust risk stratification tools.

Taken together, these studies suggest that baseline assessment using AI algorithms may improve selection for AS by enhancing non-invasive identification of patients with low- or favorable intermediate-risk disease.

AI for PCa progression prediction and exiting active surveillance

AI approaches have been increasingly investigated to predict disease progression in patients undergoing AS for PCa, with the aim of identifying patients at higher risk of pathological upgrading or clinical progression who may require definitive treatment rather than continued surveillance. The second major domain of our review, presented in Table 2, comprised studies that assessed AI as a tool for predicting disease progression in AS.

Table 2.

Studies that assessed AI as a tool in predicting progression.

Authors	Type of study/patients included	Objective	Variables used for model development	AI methods	Results
Nayan et al.⁶	Retrospective cohort/790	To compare ML methods vs. traditional LR (T-LR) in predicting grade progression	Age, PSAd, maximum percentage of core involvement, time between initial diagnosis and most recent biopsy	ML (SVM, RF, ANN, and LR)	All ML models outperformed T-LR (p < 0.001). F1 score: SVM (0.586), ML-LR (0.522), ANN (0.392), RF (0.376) vs. T-LR (0.182)
Arvidsson et al.²⁵	Prospective cohort/180	To evaluate an AI algorithm on serial biopsies in patients with PCa on AS	WSIs from prostate needle biopsies (870 biopsy images scanned)	DL (CNN)	AUC: 0.922 Sensitivity: 96% Specificity: 73% 63% of slides would not need pathology reading, with 0.55% risk of missing slides containing cancer.
Sushentsev et al.⁷	Retrospective proof-of- concept cohort/71	To investigate whether radiomic features were complementary to clinical features in predicting PCa progression on AS using ML	PSA, MRI-derived gland volume, PSA density, MRI-derived Likert score of tumor probability, target lesion localization (peripheral zone or transition zone), and target lesion biopsy GG	ML (KNN, LR, LDA, general linear model, SVM, and RF)	AUC: 0.75 (T2WI-derived radiomic features and clinical variables) vs. AUC: 0.61 (clinical variables only)
Sushentsev et al.⁸	Retrospective cohort/76	To develop and evaluate a time series radiomics (TSR) framework for predicting PCa progression on AS using serial MRI-derived radiomic features and serial PSAd	PSA, PSAd, mpMRI, bpMRI (follow-up)	Long-short-term memory (LSTM) recurrent neural network	LSTM-based model (combining TSR and serial PSAd); AUC: 0.86 This significantly outperformed a model using delta-radiomics and delta-PSAd; AUC: 0.75, p = 0.048 Its performance was comparable to expert serial MRI assessment using PRECISE; AUC: 0.84, p = 0.710
Oerther et al.⁹	Retrospective cohort/56	To evaluate whether a clinically certified AI-driven lesion detection algorithm based on bpMRI can detect csPCa in patients under AS	bpMRI (T2WI and DWI)	DL (convolutional neural network)	Algorithm performance: - for detection of ISUP ⩾1; sensitivity: 92.5%, specificity: 31% - for detection of ISUP ⩾2; sensitivity: 96.4%, specificity: 25%
Lee et al.¹⁰	Retrospective cohort/585	To develop a model incorporating diverse new observations during AS to continuously modify and update an individual’s risk of progression	At diagnosis: PSA, PSAd, prostate volume, and core positivity Likert score, histological GG, and family history In follow-up: PSA, PSAd, MRI PRECISE score, repeat biopsy information (GG and biopsy core positivity)	ML (temporal outcome-oriented clustering method) DL (recurrent neural network)	With 3 years of data collection and 3 years of follow-up, the model had a C-index of 0.79 compared with 0.70 for landmarking Cox and 0.67 for baseline Cox only. The algorithm outperformed standard statistical tools and improved its predictive power over time.
Sushentsev et al.¹¹	Retrospective cohort/64	To compare the performance of the PRECISE scoring system against MRI-derived delta-radiomics models for predicting histopathological PCa progression in patients on AS	T2WI and ADC MRI scans	ML (parenclitic networks (PN), LASSO logistic regression, and RF)	- PRECISE had the highest specificity (94.7%) and positive predictive value (90.9%) - RF had the highest sensitivity (92.6%) and negative predictive value (92.6%) - No statistically significant differences in AUC values between PRECISE and the delta-radiomics models: AUC values; PRECISE: 0.844 - PN: 0.815 (p = 0.64) - RF: 0.809 (p = 0.57) - LASSO: 0.780 (p = 0.43)
Sushentsev et al.¹²	Retrospective cohort/332	A DL-based model to predict the risk of PCa progression in patients on AS using serial PSA, adaptive PSAd (PSADA), and nonadaptive PSAd (PSADNA)	Serial PSA measurements and serial MRI scans	DL (long short-term memory recurrent neural network)	Dynamic monitoring of PSADNA consistently outperformed PSADA in predicting PCa progression on AS, both for the whole AS population and in particular for patients with baseline gland volume (BGV) <55 mL (p < 0.0001) Patients with BGV >55 mL benefit more from serial PSA monitoring.
Sun et al.¹⁵	Retrospective imaging AI study/42	To develop a DL-based method to retrospectively quantify T2 from conventional T1- and T2-weighted images	MRI (T1WI and T2WI) scans	DL (U-Net-based neural network	T2-delta values found to classify progressors from non-progressors with an AUC of 0.74
Midya et al.¹³	Retrospective cohort/121	To quantify radiomic changes on serial bpMRI among patients on AS and evaluate their association with pathologic progression on biopsy	Serial bpMRI scans, PSA, tumor volume, age	ML (maximum relevance, minimum redundancy, and RF)	The ML classifier trained using delta radiomic features showed a stronger association with biopsy upgrade compared with baseline radiomics (AUC: 0.74 vs. 0.71) and radiologist-assigned PI-RADS v2.1 scores (AUC: 0.62). Integrating delta radiomics with changes in clinical parameters (PSA and tumor volume) between baseline and follow-up resulted in the best classification model in predicting PCa progression (AUC: 0.81)
Vershinina et al.¹⁴	Retrospective cohort/76	To improve the prediction of PCa progression during AS by developing three distinct models: 1. baseline model utilizing radiomic features from initial MRI scans combined with baseline PSAd 2. a delta model incorporating feature changes between the latest and baseline available MRI scans with final PSAd 3. time series model analyzing the complete series of radiomic features and PSAd	mpMRI (baseline examination), bpMRI (follow-up), PSA, and PSAd	Traditional and advanced ML algorithms Long short-term memory for time series classification. Explainable AI (SHapley Addictive exPlanations)	- Model 1: AUC: 0.793, sensitivity: 69%, specificity: 83% - Model 2: AUC: 0.913, sensitivity: 79%, specificity: 93% - Model 3: AUC: 0.917, sensitivity: 83%, specificity: 89% Longitudinal models (models 2 and 3) outperformed the baseline model (model 1) The models showed a strong ability to distinguish progressors from non-progressors
Winzell et al.²⁶	Retrospective cohort/180 (72 on AS)	To develop a multiple-instance learning-based framework for predicting the longitudinal outcomes for PCa patients on AS	H&E-stained WSIs following a patient visit without using explicit Gleason grades	Attention-based deep multiple-instance learning and UNI-2 foundation model	Best-performing model; AUC: 0.958 However, all models showed a large drop in performance when evaluated on an external data set (AUC: 0.699)
Lowry et al.²⁷	Prospective cohort/144 (on prospective AS trial)	To assess whether U-Found, a generalized self-supervised mpMRI foundation model, offers added value in predicting histopathological progression in AS patients	The model was trained on the ADC examinations from 3244 patients	DL (self-supervised foundation model, contrastive learning, deep feature vector embeddings)	U-Found successfully detected cancer in an independent data set without being explicitly trained for lesion detection (AUC = 0.79) The prediction model incorporating a combination of clinical variables, mpMRI-derived intratumoral radiomics features, and deep embeddings generated by U-Found achieved an AUC of 0.86, outperforming models solely based on clinical and/or radiomics features

ML models based on clinical variables have demonstrated superior performance compared with traditional statistical methods. A retrospective study involving 790 patients undergoing AS showed that several ML classifiers (achieving F1 scores ranging from 0.392 to 0.586) significantly outperformed traditional logistic regression (F1 score: 0.186, p < 0.001) in predicting grade progression.⁶ Likewise, Sushentsev et al.⁷ demonstrated that adding MRI-derived radiomic features to standard clinical parameters improved baseline prediction of progression (AUC improved from 0.61 to 0.75).

Lowry et al. assessed a generalized self-supervised mpMRI foundation model to predict progression in 144 patients in a prospective AS trial. The predictive model achieved an AUC of 0.86 when combined with clinical variables and intratumoral radiomics, outperforming models that relied solely on clinical and/or radiomics features (AUC values ranging from 0.71 to 0.82).²⁷

Studies using serial MRI and delta-radiomics have also been used to predict disease progression in patients already enrolled in AS programs. These models aim to detect subtle imaging characteristics associated with biologically aggressive disease and to identify patients at higher risk of progression while on AS. DL-based methods, including long short-term recurrent neural networks, were implemented in four studies^8,9,14,15 to retrospectively detect progression in patients under AS. Sun et al.¹⁵ introduced a DL approach to quantify changes in T2 signal intensity between serial MRI scans, achieving an AUC of 0.74 for distinguishing progressors from non-progressors. MRI-derived delta-radiomics, based on ML techniques, demonstrated performance comparable to that of the PRECISE scoring system, although PRECISE showed the highest specificity of 94.7% and positive predictive value (90.9%).¹¹ In another study, Sushentsev et al.⁸ developed a time series radiomics model integrating serial MRI features and serial PSAd measurements (AUC: 0.86), which outperformed conventional delta-radiomics (AUC: 0.75) and performed comparably to expert PRECISE scoring assessment (AUC: 0.84). Similarly, Midya et al.¹³ showed that delta-radiomics from serial MRI scans in conjunction with changes in clinical parameters (PSA levels and tumor volume) can improve progression risk prediction, as AUC improved from 0.62 and 0.72, for radiologist-assigned PI-RADS scoring and baseline radiomics, respectively, to 0.74 for the ML-based delta-radiomics model. In this context, Vershinina et al.¹⁴ also demonstrated improvements from a baseline radiomics model (AUC: 0.793) to delta radiomics (AUC: 0.913) and time series models (AUC: 0.917), with the longitudinal models achieving a greater ability to distinguish progressors from non-progressors than the baseline model. In addition, convolutional neural network–based lesion detection algorithms applied to bpMRI demonstrated high sensitivity (92.5% for International Society of Urological Pathology (ISUP) 1 detection and 96.4% for ISUP 2) for detecting clinically significant disease progression during AS, although specificity remained relatively low (31.5% for ISUP 1 detection and 25% for ISUP 2).⁹

A long short-term recurrent neural network was also used to evaluate serial PSAd in predicting disease progression risk in AS patients.¹² Specifically, Sushentsev et al. found that nonadaptive PSAd outperformed adaptive PSAd and serial PSA for progression prediction, particularly in men with prostate glands below 55 mL.

Lee et al. evaluated a DL-based individualized model incorporating new clinical, radiological, and histopathological findings in AS. The model outperformed standard statistical tools (C-index of 0.79, compared to 0.70 for landmarking Cox and 0.67 for baseline Cox) and improved its predictive power over time.¹⁰ Longitudinal models operate on real-time data, enabling a more personalized approach during follow-up and supporting decision-making about whether to remain on AS or switch to radical treatment.

AI models applied to histopathological whole-slide images (WSIs) have shown particularly high predictive performance. Arvidsson et al.²⁵ developed a convolutional neural network model applied to biopsy slide images from a prospective cohort of AS patients, achieving an AUC of 0.922, with 96% sensitivity and 73% specificity for identifying patients likely to fail AS, thus reducing by almost two-thirds the number of slides requiring pathologist review, while maintaining a 0.55% risk of missing slides containing cancer. Winzell et al. used WSIs (without explicit Gleason scores) and multiple-instance learning to predict longitudinal outcomes, achieving an AUC of 0.958 with the best-performing model. However, all models showed a large drop in performance (AUC: 0.699) when evaluated on an external data set.²⁶

Overall, these studies suggest that AI models may enhance decision-making during AS through improved ability to predict disease progression and longitudinal monitoring. Notably, across the studies, models incorporating longitudinal or multimodal data consistently outperformed those based on single baseline measurements.

Discussion

ML is a discipline within AI that focuses on developing algorithms capable of representing patterns in data. Unlike classical programming, where algorithms are explicitly coded using predefined features, ML utilizes subsets of data to generate algorithms that can employ novel or distinct combinations of features and weights, often extending beyond those established by first principles.²⁹ The subset of ML techniques that use deep neural networks (DNNs) is called DL.³⁰

AS is a management strategy that involves closely monitoring a patient’s condition without giving immediate treatment, intervening only if the disease shows signs of progression.³¹ AS has been established as an acceptable management strategy for patients with low- and favorable intermediate-risk PCa. However, balancing oncological safety against the risk of overtreatment remains a challenge in urology, particularly for appropriate patient selection.^1,32 A proportion of patients initially labeled as low-risk may hide more aggressive disease that becomes apparent during follow-up. Current clinical tools included in AS protocols (PSA, mpMRI, and systematic biopsies) offer valuable information but are not always able to distinguish sufficiently indolent PCa from clinically significant PCa. In this context, AI offers the opportunity to integrate large volumes of heterogeneous clinical data to detect and identify complex patterns that are not always discernible through conventional methods.³³

In this work, we summarized the current literature on the role of AI in decision support for patients with PCa managed under AS. The results suggest that AI-based models may contribute to two critical aspects of AS management: identifying appropriate candidates for AS and predicting disease progression that requires radical treatment. Throughout the studies, ML and DL techniques were applied to a wide range of data sources, including clinical variables, imaging-derived features, and histopathological information. Overall, the results highlighted AI’s potential to improve risk stratification and personalize surveillance strategies in PCa.

When selecting patients eligible for AS, risk stratification is based on PSA levels, biopsy findings, and clinical staging, but these traditional tools may not fully capture tumor heterogeneity. Several studies in this review described ML models that integrated clinical variables, MRI-derived radiomics, biopsy-related data, or combinations thereof, improving the detection of csPCa.^5,16–24,28 In particular, imaging-based models that utilized data from bi- or multiparametric MRI scans^{18–20,22,24,28} had been proposed as a non-invasive diagnostic tool to stratify patients into distinct risk categories for PCa, thereby helping to avoid unnecessary biopsies and identify patients who might be candidates for AS.

Regarding disease progression prediction during AS, several studies have proposed AI-based models that assessed progression likelihood using clinical, imaging, and pathological data.^{6–15,25–27} Identifying patients at higher risk of progression is essential for determining when to discontinue AS to pursue definitive treatment. Notably, DL algorithms showed promising results when longitudinal data were incorporated, enabling the use of real-time data and the individualization of follow-up strategies.

Another significant aspect is the development of multimodal ML-based models that incorporate data from clinical parameters, imaging-derived features, and histopathological findings. Multimodal models significantly outperformed traditional statistical models and ML models that used only clinical or radiological information. These results suggest that the complex biological behavior of PCa is more accurately reflected when all available data are incorporated during predictive model development. As illustrated in Figure 2, by combining heterogeneous data sources and incorporating temporal changes, ML models could offer dynamic and personalized risk stratification during AS.

Figure 2.

Integration of multimodal and longitudinal data in artificial intelligence models for active surveillance decision support. Multimodal and longitudinal data allow for more dynamic and personalized risk stratification during AS.

These findings are not without their limitations. The vast majority of the included studies were based on retrospective data, often tested on single-center cohorts. Most of the proposed models lacked external validation before being applied in clinical practice. In addition, the lack of standardized AS protocols and the heterogeneity in definitions inevitably introduce bias and raise questions about ML performance across institutions. Furthermore, the rationale behind AI algorithms is usually difficult to understand, making these models “black boxes” for clinicians and thus limiting their trust.³⁴ Finally, regulatory and ethical considerations regarding algorithmic transparency and legal responsibility pose notable barriers to the adoption of these tools.³⁵ Hence, it is too early to draw definitive conclusions about the ability of AI-based predictive tools to support decision-making for AS patients.

Future research should focus on prospective multicenter studies to test and validate the efficacy of such models in real-time clinical settings, thereby enabling safer, more reliable results before incorporating them into daily urological practice.

Conclusion

AI shows significant potential to improve decision-making in AS for PCa. Across the reviewed studies, ML and DL models consistently outperformed traditional approaches, particularly when integrating multimodal data, including clinical, radiological, and histopathological features. These models demonstrated strong predictive performance for both patient selection and progression risk, supporting more accurate risk stratification and potentially reducing unnecessary biopsies and overtreatment.

AI-driven approaches, including radiomics and longitudinal modeling, further enable non-invasive and dynamic monitoring strategies tailored to individual patients. However, current evidence is predominantly based on retrospective, single-center studies with limited external validation. Issues related to model interpretability, heterogeneity in AS protocols, and regulatory considerations remain important barriers to clinical implementation.

Prospective multicenter validation and improved transparency are essential before AI-based tools can be routinely integrated into AS management.

Supplemental Material

sj-docx-1-uro-10.1177_20514158261458173 – Supplemental material for Artificial intelligence for active surveillance decision support in patients with prostate cancer: A scoping review

Supplemental material, sj-docx-1-uro-10.1177_20514158261458173 for Artificial intelligence for active surveillance decision support in patients with prostate cancer: A scoping review by Nikos E. Anthousis, Bhavan P. Rai, Charalampos Mamoulakis, Theodoros Tokas and Dimitrios Damigos in Journal of Clinical Urology

Supplemental Material

sj-docx-2-uro-10.1177_20514158261458173 – Supplemental material for Artificial intelligence for active surveillance decision support in patients with prostate cancer: A scoping review

Supplemental material, sj-docx-2-uro-10.1177_20514158261458173 for Artificial intelligence for active surveillance decision support in patients with prostate cancer: A scoping review by Nikos E. Anthousis, Bhavan P. Rai, Charalampos Mamoulakis, Theodoros Tokas and Dimitrios Damigos in Journal of Clinical Urology

Supplemental Material

sj-png-3-uro-10.1177_20514158261458173 – Supplemental material for Artificial intelligence for active surveillance decision support in patients with prostate cancer: A scoping review

Supplemental material, sj-png-3-uro-10.1177_20514158261458173 for Artificial intelligence for active surveillance decision support in patients with prostate cancer: A scoping review by Nikos E. Anthousis, Bhavan P. Rai, Charalampos Mamoulakis, Theodoros Tokas and Dimitrios Damigos in Journal of Clinical Urology

Supplemental Material

sj-pptx-4-uro-10.1177_20514158261458173 – Supplemental material for Artificial intelligence for active surveillance decision support in patients with prostate cancer: A scoping review

Supplemental material, sj-pptx-4-uro-10.1177_20514158261458173 for Artificial intelligence for active surveillance decision support in patients with prostate cancer: A scoping review by Nikos E. Anthousis, Bhavan P. Rai, Charalampos Mamoulakis, Theodoros Tokas and Dimitrios Damigos in Journal of Clinical Urology

Footnotes

Acknowledgements

None.

ORCID iDs

Nikos E. Anthousis

Bhavan P. Rai

Charalampos Mamoulakis

Theodoros Tokas

Dimitrios Damigos

Ethical considerations

Ethical approval was not required for this study, as it is a literature review based exclusively on previously published data.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Author contributions

N.A. contributed to drafting the manuscript. B.R. contributed as a co-author and reviewed the manuscript. C.M. contributed to critical authorship and editing of the manuscript. T.T. contributed to the conceptualization of the study, critical authorship, and editing of the manuscript. D.D. contributed to drafting the manuscript.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

All data supporting the findings of this study are available within the article and its supplemental materials.

Informed consent

Informed consent was not applicable, as no human participants were directly involved in the study.

Guarantor

T.T.

Supplemental material

Supplemental material for this article is available online.

References

EAU Guidelines (edn). Presented at the EAU Annual Congress. EAU Guidelines, 2026.

Walker

Marchetti

Singhal

, et al. Active surveillance for prostate cancer: selection criteria, guidelines, and outcomes. World J Urol 2022; 40(1): 35–42. https://doi.org/10.1007/s00345-021-03622-8

Obermeyer

Emanuel

EJ.

Predicting the future—big data, machine learning, and clinical medicine. N Engl J Med 2016; 375(13): 1216–1219. https://doi.org/10.1056/NEJMp1606181

Page

McKenzie

Bossuyt

, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021; 372: n71. https://doi.org/10.1136/bmj.n71

Ding

CKC

Erak

, et al. Predicting prostate cancer grade reclassification on active surveillance using a deep learning-based grading algorithm. J Natl Cancer Inst 2024; 116(10): 1683–1686.

Nayan

Salari

Bozzo

, et al. A machine learning approach to predict progression on active surveillance for prostate cancer. Urol Oncol 2022; 40(4): 161e1–161.

Sushentsev

Rundo

Blyuss

, et al. MRI-derived radiomics model for baseline prediction of prostate cancer progression on active surveillance. Sci Rep 2021; 11(1): 12917.

Sushentsev

Rundo

Abrego

, et al. Time series radiomics for the prediction of prostate cancer progression in patients on active surveillance. Eur Radiol 2023; 33(6): 3792–3800.

Oerther

Engel

Nedelcu

, et al. Prediction of upgrade to clinically significant prostate cancer in patients under active surveillance: performance of a fully automated AI-algorithm for lesion detection and classification. Prostate 2023; 83(9): 871–878.

10.

Lee

Light

Saveliev

, et al. Developing machine learning algorithms for dynamic estimation of progression during active surveillance for prostate cancer. NPJ Digit Med 2022; 5(1): 110.

11.

Sushentsev

Rundo

Blyuss

, et al. Comparative performance of MRI-derived PRECISE scores and delta-radiomics models for the prediction of prostate cancer progression in patients on active surveillance. Eur Radiol 2022; 32(1): 680–689.

12.

Sushentsev

Abrego

Colarieti

, et al. Using a recurrent neural network to inform the use of prostate-specific antigen (PSA) and PSA density for dynamic monitoring of the risk of prostate cancer progression on active surveillance. Eur Urol Open Sci 2023; 52: 36–39.

13.

Midya

Hiremath

Huber

, et al. Delta radiomic patterns on serial bi-parametric MRI are associated with pathologic upgrading in prostate cancer patients on active surveillance: preliminary findings. Front Oncol 2023; 13: 1166047.

14.

Vershinina

Sushentsev

Zaikin

, et al. Improving the potential for predicting prostate cancer progression in patients on active surveillance using explainable artificial intelligence. Cancers 2025; 17(22): 3598.

15.

Sun

Wang

Daskivich

, et al. Retrospective T2 quantification from conventional weighted MRI of the prostate based on deep learning. Front Radiol 2023; 3: 1223377.

16.

Saitta

Buffi

Avolio

, et al. Machine learning driven diagnostic pathway for clinically significant prostate cancer: the role of micro-ultrasound. World J Urol 2025; 43(1): 502.

17.

ElKarami

Deebajah

Polk

, et al. Machine learning-based prediction of upgrading on magnetic resonance imaging targeted biopsy in patients eligible for active surveillance. Urol Oncol 2022; 40(5): 191.e15–191.e20.

18.

Nicoletti

Barra

Defeudis

, et al. Virtual biopsy in prostate cancer: can machine learning distinguish low and high aggressive tumors on MRI. Annu Int Conf IEEE Eng Med Biol Soc 2021; 2021: 3374–3377.

19.

Hiremath

Shiradkar

, et al. An integrated nomogram combining deep learning, Prostate Imaging-Reporting and Data System (PI-RADS) scoring, and clinical variables for identification of clinically significant prostate cancer on biparametric MRI: a retrospective multicentre study. Lancet Digit Health 2021; 3(7): e445–e454.

20.

Liu

Zhao

Bao

, et al. Non-invasively identifying candidates of active surveillance for prostate cancer using magnetic resonance imaging radiomics. Vis Comput Ind Biomed Art 2024; 7(1): 16.

21.

Sondermann

Glaser

Rentsch

, et al. A machine learning algorithm for automatic tumour board recommendations in prostate cancer patients. BJUI Compass 2025; 6(8): e70066.

22.

Kim

Jing

Utt

, et al. An artificial intelligence model using diffusion basis spectrum imaging metrics accurately predicts clinically significant prostate cancer. J Urol 2025; 213(6): 777–785.

23.

Tsuneki

Abe

Ichihara

, et al. Inference of core needle biopsy whole slide images requiring definitive therapy for prostate cancer. BMC Cancer 2023; 23(1): 11.

24.

Liu

Fang

, et al. Multimodal nomogram combining multiparametric MRI, functional subsets of peripheral lymphocytes and PI-RADS can predict risk stratification of prostate cancer. Comput Methods Programs Biomed 2026; 273: 109086.

25.

Arvidsson

Svanemur

Marginean

, et al. Artificial intelligence for detection of prostate cancer in biopsies during active surveillance. BJU Int 2024; 134(6): 1001–1009.

26.

Winzell

Arvidsson

Åström

, et al. Longitudinal outcome prediction of prostate cancer patients on active surveillance using multiple instance learning. J Med Imaging 2025; 12(6): 061408.

27.

Lowry

Breto

Wallaengen

, et al. MRI-based patient selection for active surveillance in prostate cancer using U-Found: a generalized deep learning model. Cancer Imaging 2026; 26(1): 26.

28.

Shahait

Usamentiaga

Tong

, et al. Periprostatic adipose tissue MRI radiomics-derived features associated with clinically significant prostate cancer. J Endourol 2023; 37(10): 1156–1161.

29.

Choi

Coyner

Kalpathy-Cramer

, et al. Introduction to machine learning, neural networks, and deep learning. Transl Vis Sci Technol 2020; 9(2): 14. https://doi.org/10.1167/tvst.9.2.14

30.

Kaluarachchi

Reis

Nanayakkara

A review of recent deep learning approaches in human-centered machine learning. Sensors 2021; 21(7): 2514. https://doi.org/10.3390/s21072514

31.

Klotz

Active surveillance for prostate cancer: patient selection and management. Curr Oncol 2013; 20(suppl. 1): S11–S17.

32.

Bul

Zhu

Valdagni

, et al. Active surveillance for low risk prostate cancer worldwide: the PRIAS study. Eur Urol 2013; 63: 597–603.

33.

Van Booven

Kuchakulla

Pai

, et al. A systematic review of artificial intelligence in prostate cancer. Res Rep Urol 2021; 13: 31–39.

34.

Shuttleworth

KMJ

. Medical artificial intelligence and the black box problem: a view based on the ethical principle of “do no harm”. Intell Med 2024; 4(1): 52–57.

35.

Agrawal

Vagha

A comprehensive review of artificial intelligence in prostate cancer care: state-of-the-art diagnostic tools and future outlook. Cureus 2024; 16(8): e66225. https://doi.org/10.7759/cureus.66225

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.01 MB

0.09 MB

0.00 MB

0.03 MB

0.01 MB