Prostate cancer (PCa) is a prevalent malignancy among men worldwide, with a rising incidence [1]. In 2025, approximately 313,780 new cases of prostate cancer and 35,770 deaths are expected in the United States alone, highlighting its ongoing impact on public health [2]. Globally, PCa continues to represent 14.1% of all cancer cases and 6.8% of cancer-related mortality, with the number of new diagnoses and deaths steadily increasing in low- and middle-income regions [3]. While radical prostatectomy (RP) and radiation therapy (RT) serve as primary treatments, a subset of intermediate- and high-risk patients still experience biochemical recurrence (BCR) following therapy, typically defined by a sustained increase in prostate-specific antigen (PSA) levels [4,5]. Patients who experience BCR carry significantly worse prognoses, with elevated risks of metastasis and mortality [1]. Therefore, early and precise identification of individuals at high risk of BCR is paramount for guiding clinical management, tailoring interventions, and improving quality-adjusted survival outcomes.

Current clinical detection of PCa relies on PSA testing, bone scans, multiparametric magnetic resonance imaging, prostate-specific membrane antigen positron emission tomography and computed tomography imaging techniques, and transrectal ultrasound–guided biopsy [5]. While PSA testing is convenient, its specificity is limited [6]; transrectal ultrasound–guided biopsy is prone to sampling errors and observer subjectivity [7,8]; and multiparametric magnetic resonance imaging and prostate-specific membrane antigen Positron Emission Tomography and Computed Tomographypositron emission tomography offer superior localization and risk stratification but remain susceptible to interreader variability and qualitative interpretation bias [9]. Critically, these modalities fail to reliably extract high-dimensional quantitative imaging features, limiting their ability to reveal latent biological information [5,10].

In recent years, artificial intelligence (AI) has emerged as a promising tool in PCa diagnosis. AI outperforms traditional methodologies in processing complex datasets, uncovering hidden nonlinear relationships, and improving predictive accuracy and stability [1,4]. Magnetic resonance imaging (MRI)–based AI models integrating radiomics and deep learning have shown promising potential in predicting BCR after prostatectomy or other treatments. For instance, Hou et al [11] reported a biopsy-free AI-aided precision MRI model in the British Journal of Cancer, demonstrating robust area under the curve (AUC) values in both internal and external validation cohorts. However, current studies are mostly retrospective and exploratory in nature, with limited sample sizes, multicenter external validation deficiency, significant intermodel variability, and overfitting concerns [12]. These MRI-based AI models are intended for key clinical use cases, including pretreatment BCR risk stratification, posttreatment surveillance triage, and decision support for salvage therapy, yet their generalizability, robustness, and clinical applicability remain to be firmly established [13,14].

Several systematic reviews and meta-analyses have explored the value of AI in predicting PCa BCR. A recent systematic review by Liu et al [1] overviewed AI applications for predicting BCR following RP, highlighting that MRI-based AI models achieved the highest median AUC–receiver operating characteristic (ROC) compared with models using pathological or clinical-pathological variables alone; however, this review did not perform meta-analytic synthesis or evaluate the influence of different treatment modalities. Furthermore, a meta-analysis by Salimi et al [5] pooled data from MRI-based radiomics models for BCR prediction and reported pooled diagnostic performance, but this study only focused on radiomics-only models and did not stratify or compare model performance between internal and external validation cohorts.

To address these critical gaps, this systematic review and meta-analysis aimed to comprehensively evaluate the diagnostic performance of MRI-based AI models for predicting BCR in PCa, with a direct comparison of performance differences between internal and external validation cohorts and quantification of overall performance and clinical feasibility. These innovations fill key research vacancies and provide robust evidence for the clinical translation of MRI-based AI tools in PCa BCR prediction.

Our meta-analysis demonstrates that MRI-based AI models showed broadly comparable diagnostic performance in internal and external validation for predicting BCR of PCa. This pattern aligns with established methodological challenges in previous research [1,4]. Overfitting risk in internal validation using institution-specific data splits [15,31] and covariate shift arising from variations in MRI protocols, scanners, and preprocessing may compromise model generalizability (eg, a decline in AUC from 0.74 to 0.60 across vendors) [18,20], and the reproducibility of radiomic features remains limited in the absence of rigorous harmonization procedures [22]. These factors may lead to overly optimistic internal performance estimates, highlighting the critical importance of external validation for the clinical translation of AI models [52].

Treatment modality appeared to contribute to variability in model performance. MRI-based AI showed broadly comparable performance between RP and RT with or without HT cohorts at the pooled level, with no meaningful differences in diagnostic performance [33,44]. Treatment-stratified estimates were heterogeneous and should be interpreted cautiously, given the wide CIs in several subgroups [53]. In radiotherapy-related cohorts, performance varied across EBRT with or without HT and brachytherapy subgroups, but these findings were accompanied by substantial uncertainty. This variability may reflect treatment-related domain shift and differences in tissue response and imaging appearance, and hormone therapy may further alter glandular morphology and cellularity, potentially affecting feature extraction and radiomics signatures (eg, fibrosis) [28,32]. In contrast, RT cohorts without concurrent HT may retain more distinct radiation-induced alterations as robust imaging features [31,44] and benefit from standardized posttreatment surveillance protocols that enhance feature reproducibility [31,37]. Brachytherapy patients, on the other hand, showed a pattern of higher sensitivity but lower specificity, likely related to inflammatory artifacts [45], further highlighting the treatment-specific impacts on model generalizability.

Interestingly, our meta-analysis demonstrates that radiomics and clinical models consistently achieve prominently higher sensitivity, yet moderately lower specificity, compared to radiomics-only models [54]. This observation was corroborated by multiple included studies: Poulakis et al [18] showed that combining MRI-derived radiomics with clinical variables (such as PSA and Gleason score) raised sensitivity from 0.81 to 0.89 but reduced specificity from 0.82 to 0.71; Jambor et al [34] and Hou et al [11] also reported that the integration of clinical risk factors and imaging features enhances the model’s ability to identify true BCR cases, but at the expense of increased FPs [11,34]. Mechanistically, this phenomenon can be attributed to the complementary information from multisource data fusion: clinical variables capture systemic and biological aspects of tumor aggressiveness not reflected in imaging, which increases detection of at-risk patients (higher sensitivity); however, many clinical high-risk markers (eg, elevated PSA and higher Gleason score) are nonspecific and may also be present in patients without recurrence, increasing the likelihood of misclassification and thus lowering specificity [34,35,45,55]. This trade-off is further supported by other included studies such as Park et al [35], Zhu et al [45], and Sun et al [56], all of which found that adding clinical variables to radiomics models improved recall but led to a rise in FP results. This trade-off between elevated sensitivity and reduced specificity highlights that optimizing the sensitivity-specificity balance is a core challenge for clinical AI model development [57]. The ideal balance is context dependent, varying with the clinical consequences of missed BCR (FNs) versus unnecessary interventions (FPs) [58]. Future model development must prioritize context-specific optimization of this balance to maximize practical clinical value [59].

In our meta-analysis, subgroup analysis based on the timing of MRI image acquisition revealed that models using pretreatment MRI exhibited improved predictive specificity for BCR compared to those using posttreatment MRI [60]. This finding is consistent with prior research, which suggests that pretreatment MRI provides more reliable information on tumor characteristics, as it is unaffected by posttreatment changes, such as fibrosis, necrosis, or tissue remodeling [61]. In contrast, posttreatment MRI may be influenced by these treatment-induced alterations, which can mimic tumor recurrence and confound image interpretation [62]. Such nontumor-related changes are more likely to appear on posttreatment scans, potentially leading to misinterpretation and FPs, thereby reducing specificity [63]. Therefore, careful consideration of MRI timing is critical for optimizing the clinical utility of MRI-based AI models for BCR prediction, particularly for enhancing specificity and minimizing FPs [64].

A recent systematic review by Liu et al [1] provided a valuable overview of AI applications for predicting BCR following RP, highlighting that MRI-based AI models achieved the highest median AUC-ROC (0.90) compared to models using pathological or clinicopathological variables alone. While Liu et al [65] offered important insights, their review focused specifically on postprostatectomy cohorts and did not perform meta-analytic synthesis or examine the influence of different treatment modalities. Our study builds upon this foundation by being the first to quantitatively synthesize the diagnostic performance of MRI-based AI models for BCR prediction across both RP and RT populations [11]. In addition, we directly evaluate the impact of treatment type on model accuracy.

Furthermore, a meta-analysis by Salimi et al [5] specifically pooled data from 24 studies of MRI-based radiomics models for BCR prediction, reporting pooled sensitivity, specificity, and AUC of 0.72, 0.78, and 0.75 for radiomics-only models, with clinical-radiomics models reaching an AUC of 0.88. While our pooled estimates align with Salimi et al, our meta-analysis introduces key methodological advances. Specifically, we are the first to stratify and compare model performance between internal and external validation cohorts. This novel analysis provides crucial quantitative evidence on model generalizability—essential for clinical translation [66]. Our findings show robust performance in internal validation and a slightly lower, but not statistically significant, AUC in external settings. This highlights the critical need for multicenter data harmonization and rigorous external testing [67]. In summary, these enhancements make our meta-analysis the first to comprehensively evaluate MRI-based AI for BCR prediction across diverse treatments and validation paradigms, offering new insights for future research and clinical use [11].

Our analysis identified substantial heterogeneity across studies that may affect the pooled diagnostic performance. Meta-regression suggested that AI method, AI model type, timing of MRI acquisition, and treatment modality were potential contributors to heterogeneity [68]. Several studies in both internal and external validation cohorts behaved as potential outliers and may have contributed to between-study variability. In addition to these identified factors, other unmeasured variables, such as differences in image preprocessing strategies and modeling approaches, may also have contributed to the observed heterogeneity [22]. These findings underscore the necessity of enhanced methodological standardization to improve model generalizability [20]. To address this heterogeneity, unified data protocols and reporting standards are imperative: uniform MRI acquisition and preprocessing protocols should be implemented to reduce equipment-related variability; diagnostic contingency tables (TP, FP, FN, or TN) should be explicitly reported to eliminate back-calculation and improve transparency; and consensus-based BCR definition guidelines should be adopted to reduce diagnostic-criterion heterogeneity across studies [69].

This meta-analysis has several notable limitations that may affect the interpretation and generalizability of its findings. First, most included studies were retrospective, with very few prospective studies included [27,30]. Retrospective designs inherently carry a higher risk of selection and information bias, particularly in the context of AI model training where unmeasured confounding and dataset imbalance may compromise predictive validity [70]. Second, many studies involved small sample sizes in either the training or validation cohorts [71]. Such limited datasets increase the risk of overfitting and reduce statistical robustness, potentially inflating diagnostic performance metrics [72]. Future studies should incorporate larger, balanced datasets with independent validation to improve model robustness [73]. Third, the definition of BCR was not uniform across all included studies. Although the AUA and Phoenix criteria were most commonly used, some studies used alternative or study-specific definitions, which may introduce inconsistency in diagnostic labeling. Although our subgroup analyses did not suggest meaningful performance differences between different reference standards, the lack of standardization remains a concern and highlights the need for consensus-based guidelines in future research [74]. Fourth, although the Deeks’ funnel plot did not indicate small-study effects, the limited number of included studies reduced the statistical power of this test, and the possibility of bias cannot be entirely excluded [75]. Fifth, selecting the best-performing model from each study may introduce optimistic bias. By focusing on top-performing models in each individual study, the performance metrics may be overly optimistic and not representative of real-world clinical outcomes [76]. In addition to overfitting, another significant limitation of the included AI models is their limited interpretability and transparency, which impedes clinical trust and regulatory acceptance [77]. Finally, and critically, due to lack of data, this study could not compare AI performance against MRI radiologists. This head-to-head comparison is indispensable for determining whether AI systems perform better than, equivalent to, or worse than human experts, serving as a key prerequisite for their clinical implementation [78]. Without this critical human benchmark, it remains unclear whether these MRI-based AI models offer genuine incremental clinical value or merely replicate the existing diagnostic capabilities of radiologists [79]. Prior systematic reviews have highlighted that while AI models sometimes report diagnostic accuracy comparable to or exceeding that of radiologists, these results often lack consistency when evaluated on independent datasets or in real-world settings [80]. As radiologist-level benchmarking is crucial for clinical adoption, future studies should directly compare AI tools against radiologists across different levels of expertise [81].

This meta-analysis innovatively quantifies and compares MRI-based AI model performance for BCR prediction between internal and external validation cohorts—a core clinical translation issue understudied in prior research, and our findings confirm these models exhibit promising diagnostic performance in both cohorts [82]. Distinct from previous reviews limited to single treatment cohorts or radiomics-only models, this study comprehensively evaluates AI performance across all major PCa treatment modalities and integrates machine learning and deep learning approaches, with subgroup findings revealing that AI models perform particularly well in patients who received EBRT alone [83]. These insights may inform personalized follow-up protocols based on treatment modality and risk stratification. In terms of clinical practicality, AI could augment radiologists’ ability to interpret complex MRI data, especially by extracting subtle, high-dimensional imaging features beyond the perceptual threshold of human observers. This capability may lead to earlier detection of recurrence risks and improved decision-making around salvage therapy [10]. Nevertheless, caution is warranted. PCa management involves multifactorial decision-making that integrates clinical context, patient preferences, comorbidities, and functional outcomes—areas where human judgment remains indispensable [80]. Therefore, AI should be positioned as a support system rather than a replacement for human expertise. Importantly, while algorithmic refinement and model optimization remain active areas of research, future gains in clinical applicability may depend more critically on improving the precision and consistency of imaging data itself [84]. Variability in image acquisition protocols—including differences in scanner hardware, field strength, pulse sequences, and operator technique—can significantly affect the quality of input data and, by extension, the reliability of AI outputs [85,86]. Enhancing the standardization of MRI protocols and ensuring high-fidelity imaging across institutions may serve as a more foundational step toward improving model robustness and generalizability [87]. Finally, the future of AI in PCa imaging will likely depend on its ability to synthesize multimodal data—including radiologic, genomics, and clinical biomarkers—into cohesive, interpretable frameworks [88]. Advances in federated learning and privacy-preserving technologies may also help overcome data-sharing barriers and facilitate collaborative model development across institutions [89,90]. Addressing these technical and regulatory challenges is essential for translating AI innovations into improved patient outcomes and more efficient PCa care pathways [91].

In conclusion, this meta-analysis contributes by quantitatively comparing the generalizability of MRI-based AI models through a synthesis of their performance in internal versus external validation for BCR prediction and further examines the influence of different treatment modalities on model performance relative to earlier reviews. Treatment-stratified analyses suggested heterogeneous performance across modalities, particularly within radiotherapy-related subgroups; however, these findings should be interpreted as hypothesis generating, given limited study numbers and wide CIs. Pretreatment MRI was identified to yield superior diagnostic specificity for BCR prediction, providing actionable guidance for future AI model development and clinical imaging workflow design. Collectively, these findings indicate that MRI-based AI models are promising investigational tools for PCa BCR prediction but are not yet ready for routine clinical use, due to the performance variation in external validation, the predominance of retrospective evidence, and unresolved methodological limitations. These tools may serve as adjunctive support for clinical decision-making in specific research settings, and their broader clinical translation requires large, prospective multicenter studies with harmonized imaging standards, rigorous external validation, and improved model interpretability.