Decades of AI in Prostate Cancer Show Technical Gains but Clinical Utility Remains Elusive

For three decades, artificial intelligence has been applied to prostate cancer (PCa), yet definitive evidence that these sophisticated tools improve long-term patient outcomes remains limited. This narrative review, tracing the field from 1994 through 2025, reveals a recurring pattern: while technical performance has achieved impressive levels, particularly in retrospective analyses, the translation into consistent clinical benefit has been slow and fraught with systemic barriers. The field has transitioned from early shallow neural networks relying on hand-crafted features to modern deep learning and large language models, but the journey from algorithmic promise to real-world clinical utility depends heavily on workflow integration, data standardization, and regulatory acceptance.

Prostate cancer is a malignancy frequently diagnosed in men, and its management is inherently complex, involving decisions that range from Gleason grading—which varies among pathologists—to screening protocols and treatment choices based on individual patient values. Since the advent of neural network studies in 1994, AI has been investigated across nearly every major decision point in PCa. Despite the rapid technical evolution, no prospective study by the end of 2025 had demonstrated an improvement in overall survival or quality of life for patients with localized PCa. This absence of long-term data is not necessarily a failure of the technology, but rather a reflection of the long natural history of localized PCa, which necessitates extended follow-up periods to observe survival or quality-of-life effects. The current evidence suggests that the lack of immediate long-term data should not be interpreted as a lack of clinical utility.

The initial phase of AI in PCa, spanning 1994 to 2016, was characterized by the application of machine learning (ML) using shallow architectures like artificial neural networks (ANN), support vector machines, and Bayesian classifiers. In this era, ML offered no consistent advantage over traditional statistics when applied to low-dimensional clinical variables. Its practical value was most evident when models utilized high-dimensional molecular or image data. Early studies, such as one from 1994 using PSA, digital rectal examination, and transrectal ultrasound data, reported up to 87% accuracy. Furthermore, early ANNs often operated within the same feature space as logistic regression and lacked predictive explainability, preventing widespread clinical establishment.

The field saw a shift in focus from low-dimensional clinical data to image-based AI, particularly in pathology and imaging. Between 2007 and 2015, the constraint shifted to feature design, as AI relied heavily on hand-crafted features to encode institutional characteristics like staining protocols or scanner hardware, rather than just disease biology. Tabesh et al. set a benchmark in 2007 using an SVM classifier, achieving 96.7% accuracy for cancer detection. In imaging, ANN-based TRUS systems began extracting diagnostic information from raw signal data, and multiparametric MRI CAD systems reported AUC values between 0.83 and 0.89. However, cross-institutional generalizability remained unestablished because these systems were tied to specific institutional characteristics. A notable exception was the Stockholm3 trial, which utilized a model integrating six protein biomarkers, 232 SNPs, and clinical variables to detect Gleason score 7 or higher cancer in over 47,000 men. This model reduced biopsies by 32% and entered routine use in Scandinavia, demonstrating the critical role of clinical fit in adoption, even if the underlying technology used simpler logistic regression.

The transition to deep learning around 2016 marked a significant technical leap. Convolutional neural networks (CNN) allowed researchers to apply AI to prostate histopathology without requiring manual feature engineering. The CAMELYON16 challenge demonstrated that deep learning could match pathologist-level cancer detection. While this reduced the burden of manual feature design, it introduced new barriers: predictions became less transparent than those of shallow ANNs, and sensitivity to acquisition differences persisted due to the use of raw pixel data. In pathology, deep-learning systems reported tumor-detection AUC values above 0.980 between 2019 and 2020, often outperforming general pathologists. However, detection and grading did not perform equally; for example, one system achieved an AUC of 0.997 for tumor detection but only a kappa of 0.62 for Gleason grading, which was at the lower bound of expert performance. This highlighted the distinction between high-level detection and nuanced pathological grading.

As technical performance approached a ceiling in detection (AUC > 0.99 in pathology), the focus shifted toward treatment guidance and population-level validation. By 2021, regulatory milestones emerged, such as the US Food and Drug Administration authorizing Paige Prostate for AI-assisted cancer detection on digitized biopsy slides. This deployment milestone saw pathologists assisted by the system increase sensitivity from 88.7% to 96.6%. Concurrently, research moved beyond single-center studies to address the generalization gap. The PANDA challenge dataset and multi-institutional studies showed high performance across varied data, yet persistent differences remained between near-perfect tumor detection and variable Gleason grading across institutions. In MRI, algorithms often performed less reliably on external cohorts due to variations in scanners and acquisition protocols, suggesting that standardized inputs are crucial for easing evaluation.

The most recent developments, spanning 2024 to 2025, have seen the entry of large language models (LLMs) and agentic AI into the clinical discourse. ChatGPT’s release in late 2022 accelerated interest in these tools. In clinical workflows, LLMs have been used for patient education and communication, with some studies showing positive results in reducing clinician workload and improving patient confidence scores. However, these applications currently support supervised decision support rather than autonomous care, as LLMs can produce responses that are incomplete or inaccurate. Furthermore, complex agentic systems—which combine evidence retrieval, tool use, and multistep workflows—have shown high rates of correct clinical conclusions in proof-of-concept studies, but their citations and claims require rigorous source verification, underscoring the need for human oversight and traceable evidence.

While survival and quality-of-life benefits remain elusive in prospective trials, several clinically meaningful intermediate outcomes have been reported following clinical integration. In a study of 82 patients, AI deployment reduced the use of immunohistochemistry per detected PCa case (relative risk 0.55). Another randomized trial involving AI-guided cognitive targeted biopsy and systematic biopsy increased the intention-to-treat patient-level detection of clinically significant cancer from 46.56% to 58.64%. Furthermore, AI-augmented surgical guidance showed positive surgical margins of 22% versus 39% with standard guidance. These process outcomes and early functional endpoints are vital, but they must be evaluated separately from the long-term survival question, which, given the decade-long course of localized PCa, cannot yet be answered by current AI trials.

Looking ahead, the field is grappling with persistent challenges across three decades, including safety, explainability, fairness, regulation, and translation. Safety risks escalate with each AI generation, ranging from hypothetical risks in early ANNs to model drift and flawed reasoning in LLMs. Explainability remains a hurdle, as post-hoc explanations can be misleading if the model’s rationale does not map to established morphological criteria. Fairness is a growing concern, with recent empirical studies indicating attenuated prognostic signals for African American patients in multimodal AI models, necessitating multiethnic validation and corrective strategies like synthetic data generation. Regulatory bodies, including the EU AI Act, are attempting to establish frameworks, but these often lag behind technological advancements, creating a gap between rapid innovation and safe clinical deployment.

Subject of Research: The evolution and current limitations of artificial intelligence applications in prostate cancer diagnosis and treatment.

Article Title: Three decades of artificial intelligence in prostate cancer show technical progress while evidence of clinical benefit is still emerging

Article References: Yue, W., & Padanilam, B. J. (2026). Three decades of artificial intelligence in prostate cancer show technical progress while evidence of clinical benefit is still emerging. Discover Artificial Intelligence, 6(1), Article 1659. https://doi.org/10.1007/s44163-026-02605-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02605-1

Keywords: Prostate Cancer, Artificial Intelligence, Deep Learning, Medical Imaging, Clinical Translation, Machine Learning, AI Ethics, Healthcare Technology, Three, decades, artificial, intelligence