From prostate-specific antigen to precision: The future of prostate cancer diagnosis with artificial intelligence, biomarkers, and imaging.

1. Introduction
Currently, prostate cancer (PCa) screening, diagnosis, and prognosis without prostate-specific antigen (PSA) are almost impossible. Until PSA was isolated in 1979 by Dr. Ming Chu’s research group[1,2] and later, the development of an enzyme-linked immunosorbent assay immunoassay that could be used for blood detection,[3] one of the only biomarkers available to evaluate PCa, prostatic acid phosphatase, was limited in men with metastatic bone disease.[4] From these early PSA studies, it was not believed that PSA could be used as a screening tool for PCa; until the early 1990s, it was primarily used for monitoring treatment response in patients who already had an established PCa diagnosis,[2,5] and it was already known that patients with other prostatic diseases, such as benign prostatic hyperplasia (BPH) and prostatitis, could present higher PSA levels. The use of PSA as a first-line screening blood test comes mainly from Catalona’s 1991[6,7] and 1994[7] trials, which compared digital rectal examination with PSA testing for the early detection of PCa, leading to its approval as an aid for early detection by the US Food and Drug Administration. From this time on, a new era started, marked by widespread PSA testing, with initial results of a decline of 45% to 70% in PCa mortality by the year 2000 attributed to screening.[8]
Over time, some limitations of the widespread screening for PCa using PSA blood testing have raised concerns. Although PSA has a considerably high sensitivity of 89% for the detection of early PCa, at a consensual biopsy threshold of 4 ng/mL, its specificity reaches 54% with a positive predictive value of 20% to 30%.[9] This low specificity occurs because PSA is also expressed in benign prostate conditions, which leads to unnecessary biopsies.[9,10] Another disadvantage of widespread PSA screening is the risk of overdiagnosis and subsequent overtreatment. Prostate cancer is heterogeneous; it can vary from indolent to aggressive, and patients can receive a cancer diagnosis without survival benefits from this detection and even suffer more harm from interventions, monitoring, and treatment.
Considering the interindividual and intratumoral heterogeneity of features in PCa, we face underwhelming pretreatment assessments and risk stratifications, leading to worse treatment plans. In this context, there is an urgent need for more precise diagnostic tools that incorporate the latest technologies available in the biomedical field, such as artificial intelligence (AI), novel biomarkers, and advanced imaging.
This study explored these emerging trends and the integration of advanced technologies into diagnostic techniques, which will shape the future of PCa diagnosis.

2. Limitations of prostate-specific antigen–based diagnosis
Prostate cancer can be detected through biopsy, even at normal PSA levels.[11] Moreover, increases in PCa detection rate are proportional to increases in PSA’s thresholds for biopsy.[12] The common threshold of 4 ng/mL was consensually determined in the 1990s as a tradeoff between sensitivity and specificity.[13] No cutoff point of PSA value shows high sensitivity and high specificity for PCa simultaneously; the reality is that the risk for PCa exists in all PSA values.[14] Multiple studies have evaluated the sensitivity and specificity of PSA for diagnosing PCa, yielding varying results depending on factors such as age, intrinsic population characteristics, and the presence of PCa symptoms before testing. The 4.1 ng/mL threshold in one study showed a 6.2% false-positive rate, with a 20.5% sensitivity.[14] The same study analyzed lowering this cutoff to 1.1 ng/mL, increasing the detection rate of cancer cases to 83.4% but subjecting 61.1% of men without cancer to prostate biopsy. In comparison, the recommended threshold of 2.6 ng/mL showed a sensitivity of 40.5%. The sensitivity of PSA also tended to increase among men with a Gleason score of 8 or higher, reaching a 50.9% value at 4.1 ng/mL.[14]
In this scenario of doubt between the advantages and disadvantages of lowering or increasing PSA cutoff levels for biopsy, the risks of unnecessary biopsies and overtreatment should be considered. Although prostate biopsy techniques have improved over time, it is important to note that this procedure is not completely indolent, considering the risk of rectal bleeding, hematospermia, hematuria, urinary tract infection, and pain, with a small number of false-negative results.[15–18] In addition to the risks inherent to biopsy, there are also aspects of overdiagnosis and overtreatment associated with PSA screening. Low-grade and low-volume tumors can be monitored and treated only with evidence of growth or changes in aggressiveness, which avoids unnecessary treatment and maintains low rates of metastasis and cancer-related mortality.[19,20] Data from the CAPSURE PCa registry[20] also show that 92% to 98% of the patients who had the lowest tumor scores, who were also presumably eligible for active surveillance, were treated with radical surgery, radiation, or hormone therapy, which means that at least 92% of men who had low tumor scores and a 10-year risk of PCa death of 2.8% received aggressive treatment.[20] The quality of life should also be considered when evaluating the effects of screening and overtreatment programs. An analysis of the European Randomized Study of Screening for Prostate Cancer[21] concluded that the benefit of PSA screening was diminished by the loss of quality-adjusted life-years due to the postdiagnosis long-term effects of cancer treatment, such as sexual dysfunction, poor urinary function, and compromised bowel function.[22–24]
If overdiagnosis and overtreatment pose risks to patients, and not diagnosing aggressive PCa soon enough results in higher mortality, we still face other challenges, such as distinguishing between indolent and aggressive PCa. For example, epidemiological studies frequently use a combination of TNM stage, Gleason score, and PSA value at diagnosis to define tumor aggressiveness, mainly because these data are available in population-based studies. However, the variability in outcome definitions between different studies limits the possibility of comparing and combining the results of meta-analyses. Although we currently have a better understanding of the nuanced differences between indolent and aggressive PCa, there is still uncertainty in accurately predicting their behavior, which can lead to treating an indolent tumor too aggressively or undertreating an aggressive tumor.
In this context of uncertainty and the lack of a homogeneous definition, the scientific community has made many efforts to develop better diagnostic pathways, not by excluding PSA, but by finding solutions that can enhance its use.

3. Artificial intelligence and machine learning in prostate cancer diagnosis
Artificial intelligence refers to the ability of machines and software to perform tasks that require human intellect. Machine learning (ML) is a branch of AI focused on developing models and algorithms that can improve performance by exposing and identifying patterns in data over time without having been explicitly programmed for every act or decision. Considering this, there are many applicable situations for AI and ML diagnoses. Because ML is especially useful in image recognition and the identification of image patterns, the 2 fields in which it is being primarily researched are radiology and, more specifically, magnetic resonance imaging (MRI) interpretation and pathology diagnosis.

3.1. Artificial intelligence in imaging
Multiparametric MRI (mpMRI) integrates functional and anatomical MRI sequences. Its main components are T1-weighted, T2-weighted, and diffusion-weighted images, including apparent diffusion coefficient maps and dynamic contrast-enhanced images.
The positive predictive value of mpMRI for suspicious clinically significant PCa reached 40% in a meta-analysis, a value that is highly dependent on disease prevalence,[25] whereas its negative predictive value reached 90.8%.[26] Multiparametric MRI is also helpful for prognostic measurements, considering the functional sequences that aid in predicting tumor behavior. However, variations in image acquisition methods and protocols between institutions can result in inconsistencies in imaging quality, making image comparisons difficult. In addition, mpMRI interpretation requires a steep learning curve while still being subject to variability in interobserver interpretation.[27] In this context, AI applications, specifically ML, in prostate mpMRI have been developed and enhanced for accurate image interpretation, specifically for PCa lesion detection, volume estimation, and characterization. A summary of recent studies evaluating artificial intelligence and machine learning applications in prostate imaging is presented in Table 1. It is also known that computer-assisted diagnosis increases the sensitivity and specificity of the detection of PCa on mpMRI, complementing radiologists’ assessments. One of the most important evaluations in PCa imaging is prostate segmentation, which is the process of identifying and delineating the boundaries of the gland in images to accurately separate the prostate from the surrounding tissues and identify its deformable capsule. Prostate segmentation has multiple applications, ranging from cancer diagnosis and surgical planning to prostate fusion biopsy and brachytherapy.[28] Prostate segmentation has been one of the main focuses of AI research on mpMRI, such as the PROMISE12 challenge dataset, which achieved very good results on AI-tooled prostate segmentation.[29] In terms of prostate volume estimation, which is a highly valuable measure for the assessment of BPH treatment, surgical planning of prostatectomy, and PCa prognosis, there are good results from AI analysis of mpMRI and studies that combine volume estimation and segmentation.[30]
Artificial intelligence can also assist in lesion identification, tumor volume estimation, and lesion segmentation. One of the biggest challenges in these processes was creating a fully automated system; until then,[31] all regions of interest had to be manually located before and after the Deep Learning (DL) model had been trained. In 2022, a model that simultaneously performs fully automated detection, segmentation, and Gleason Grade estimation using mpMRI was developed with promising results.[32]
To standardize the acquisition, interpretation, and reporting of prostate mpMRI, the Prostate Imaging Reporting and Data System (PI-RADS) was introduced in 2012 by the European Society of Urogenital Radiology, with subsequent updates until the current version of 2019, PI-RADS v2.1.[33] Prostate Imaging Reporting and Data System evaluates individual imaging sequences, scoring their findings into a risk category classified from 1 to 5, with an increasing likelihood of significant PCa. The PI-RADS depends on mpMRI interpretation and the challenges already known for reducing the heterogeneity among MRI readers, the varying radiologist experiences, and the time required to interpret the images. Machine learning techniques offer solutions for accurate and quick classification. Multiple studies have shown that AI algorithms can provide PI-RADS scores for prostate lesions and enhance the performance.[36]
Radiomics is the process of computationally extracting features from radiographic images with the goal of characterizing disease patterns, and its use has been researched with promising results in risk stratification,[34] prediction of the Gleason Score of a lesion,[35] and treatment outcomes.[36] Regarding risk stratification, a multicenter study demonstrated the role of peritumoral radiomics associated with heterogeneity patterns around the tumor as a predictor of PCa risk based on biparametric MRI.[34]

3.2. Artificial intelligence in pathology
Digital pathology involves the conversion of traditional glass microscope slides into high-resolution digital images using specialized scanning devices. These whole-slide images, which are digitally converted from original slides, are particularly useful for applying AI algorithms to identify areas of concern, confirm initial diagnoses, and enhance diagnostic accuracy. In addition, AI algorithms can provide secondary opinions for evaluation by pathologists. One example of an AI-enhanced pathology analysis is DeepGleason.[41] This tool utilizes deep neural networks to automatically grade PCa via digital pathology, thereby enhancing the consistency and accuracy of Gleason grading by providing secondary opinions to pathologists. A meta-analysis of AI diagnostic accuracy for PCa[42] found a sensitivity of 77% to 87%, a specificity of 82% to 90% for AI-driven Gleason grading, and an accuracy of 83.7% to 98.3% in differentiating PCa from BPH with AI (Table 2).
For patients who already have histopathology performed, AI can assist clinicians in distinguishing tumors that require definitive treatment from those suitable for active surveillance by integrating additional molecular and microenvironmental biomarkers with conventional pathology and imaging. Practically, this means combining digitally quantified histomorphology (automated Gleason pattern quantification, percent pattern 4, detection of cribriform and intraductal components,[43] tumor volume, perineural invasion) with molecular assays and multiomic readouts—for example, tissue-based genomic classifiers[44] (Decipher, Oncotype DX, and Prolaris), targeted gene expression signatures and copy-number/mutation profiles[44] (e.g., TP53, PTEN, BRCA alterations), DNA-methylation signatures,[45] urinary/exosomal markers[46,47] (PCA3, ExoDx/SelectMDx), proteomic signatures,[48] and spatial transcriptomic measures of tumor–stromal and immune contexture.[49]
However, the limitations of AI in pathology should be acknowledged.[50] Artificial intelligence models can be vulnerable to sampling bias, tumor heterogeneity (a biopsy may miss a small high-grade focus), image variability (staining/scanner differences), training bias, a lack of external generalizability, limited prospective validation in randomized settings, and issues of interpretability and regulatory certification. These constraints create real risks, such as missed high-grade tumors, over- or undertreatment from spurious predictions, and reduced performance, when models are deployed in different laboratories or populations.

3.3. Artificial intelligence in risk-prediction models
The use of AI in risk-prediction models has been widely explored by combining simple clinical data with biomarkers to analyze more complex and advanced data. An ML model that combined PSA, free PSA, and age for improved diagnostic accuracy in screening for PCa had an area under the curve (AUC) of 0.72 versus 0.63 for PSA alone.[56] Another relevant topic is the integration of AI with clinical and genomic profiling, which enables a more precise risk stratification process and, consequently, more personalized management of PCa. Multiomics involves integrating data from different omics layers (including transcriptomics, genomics, and proteomics) to better understand biological systems.[57] One way in which this integration can work is to employ AI algorithms to develop novel biomarkers from multiomics using a combination of clinical data, digital pathology, and genomic information. Studies that used AI models to analyze genomic and clinical data could predict which patients with PCa were more likely to benefit from certain treatments.[58] A recent analysis of large-scale multiomics datasets that included data from The Cancer Genome Atlas used an integrative clustering approach to determine distinct molecular subtypes associated with potential prognostic biomarkers in PCa; among these, CCNB1, FOXM1, and RAD51 emerged as promising candidates for prognostic evaluation.[59] Artificial intelligence tools can also automate the processing and analysis of large genomic datasets, allowing the incorporation of genomic profiling into everyday clinical care. As an example, the Oncotype DX assay,[60] which uses prostate biopsy needle samples to predict tumor aggressiveness and measure RNA expression of PCa-related genes, generates a Genomic Prostate Score with an increasing likelihood of adverse pathology (Gleason pattern 4–5, non–organ-confined disease). Another recent study evaluated an AI prediction model for the passive screening of patients with PCa by analyzing data from electronic medical records and identifying high-risk patients, achieving equivalent sensitivity to PSA and 38% higher specificity.[52] A prospective cohort of nearly 10,000 patients with localized PCa was analyzed for risk stratification using a genomic classifier integrated with digital pathology AI to verify the satisfactory performance of these tools beyond a clinical-genomic model.
It is also important to note the revolution in AI caused by large language models such as ChatGPT, which can be applied to PCa diagnosis primarily by leveraging their ability to understand, integrate, and generate insights from large amounts of complex, unstructured data. A recent prospective study[61] that included 500 patients who used the GPT-4 to extract clinical variables from unstructured reports combined with rule-based algorithms classified the national comprehensive cancer network risk groups. In 358 patients, GPT-4 with retrieval-augmented generation produced more accurate treatment plans (64% fully correct, 36% partially correct) than GPT-4 alone (35% fully correct, 65% partially correct), and retrieval-augmented generation significantly reduced the treatment suggestions by GPT-4 (32% vs. 71% for GPT-4 alone).

4. Integrating artificial intelligence, biomarkers, and imaging for a unified diagnostic model
“Multimodal precision diagnosis” of PCa utilizes AI to integrate multiple data sources, including genetic and molecular biomarkers, imaging, and clinical data, to provide accurate and personalized diagnosis and risk assessment of the disease. Figure 1 summarizes the current landscape of prostate cancer diagnosis, including artificial intelligence in imaging and pathology, novel biomarkers, multimodal integration, and regulatory considerations. Many studies have been conducted in this regard, with the results showing that multimodal AI (integrating DL lesion suspicion levels, PSA, prostate volume, patient age, and MRI-based lesion volumes) can outperform both clinical and MRI-only AI in the detection of clinically significant PCa.[54] Multimodal tools also improve risk stratification of suspected lesions and reduce the risk of unnecessary biopsies.[62]
It is important to highlight that the combination of PSA, MRI, and biopsy cannot be substituted for AI alone. Prostate-specific antigen testing is usually performed before obtaining tissue samples; specifically, AI-guided tools can play a valuable role in reducing unnecessary biopsies by combining PSA measurements with multiparametric MRI data, demographic characteristics such as age and ethnicity, clinical symptoms, and complementary laboratory findings. By integrating these diverse inputs, AI models can more accurately identify patients at an elevated risk of clinically significant PCa, thereby improving diagnostic precision while limiting avoidable procedures. A recent study analyzed a DL model for upfront risk stratification of clinically significant PCa and the reduction of unnecessary biopsies using clinical data, PSA levels, MRI, and biopsy results, with an AUC of 0.822 and an ability to reduce unnecessary biopsies by 43.4%.[63]
However, these models face challenges in their implementation in the daily practice of urologists. The factors to be considered when analyzing the values of implementing a new AI multimodal system include the significant cost of acquiring personalized AI software for each institution and its maintenance over time, the resources required to convert conventional pathology into a fully digital format, and the training of physicians on these new technologies; these can be substantially overwhelming, mainly for smaller institutions.[64]
Another barrier to the widespread use of AI in PCa diagnostic settings is the need to standardize the diagnostic methods. Magnetic resonance imaging acquisition and interpretation suffer the most from a lack of standardization among institutions and research studies. It is highly important that institutions and researchers use the same format for digital mpMRI images and equivalent/reproducible software packages for analysis because the heterogeneity in currently available data limits the possibility of comparing results among multiple centers.
To achieve the universal application of AI in clinical settings, it is imperative to have regulatory measures for this new technology. Artificial intelligence technologies should adhere to regulatory standards regarding safety, accuracy, and effectiveness, not only for quality control but also for consistent progress in the field. Another important aspect is the need for transparency in how AI works for its users, so that it is accessible and user-friendly for everyday practitioners, developers, and those who will inspect and regulate it. It is also important to note that AI technology providers should be accountable for their actions through the application of supervision by healthcare providers and by having all parties involved in the development and deployment of AI accountable for errors and harm to individuals that could occur. In this context, some regulatory organs have already introduced actions to better address the risks of the inadvertent use of AI, such as the Artificial Intelligence Act (AIA) by the European Commission.[65]

5. Future perspectives and clinical implementation
Considering the expansion of AI in the medical field, these tools have the potential to revolutionize urological practice by providing better diagnostic accuracy, improving treatment planning, reducing unnecessary biopsies, and allowing better patient outcomes.
To achieve the broad clinical adoption of AI, it is imperative to establish clear regulatory frameworks tailored to this technology. Artificial intelligence tools should comply with defined standards for safety, accuracy, and clinical effectiveness, such as those outlined by the US Food and Drug Administration Software as a Medical Device guidelines or the European Commission’s AIA, to ensure both quality control and reproducible progress. Transparency is also critical; AI systems should provide interpretable outputs and clear explanations of their decision-making processes, enabling clinicians, regulators, and developers to understand, validate, and trust the results.[65] Accountability mechanisms must be formalized, including postmarket monitoring, error reporting, and liability structures, to hold developers, healthcare institutions, and AI providers responsible for harm or misdiagnoses. Several regulatory bodies have begun implementing such measures. For instance, the AIA classifies AI systems into risk-based categories with specific requirements for high-risk medical applications, including mandatory conformity assessments and ongoing performance monitoring.
Considering both the pros and cons of the widespread use of AI tools in urology, future research on the topic must focus on large-scale validation studies and the real-world implementation of this technology, so that biases regarding AI algorithms being “trained” only on data from a specific demographic or poorly standardized images can be reduced. It is important that AI algorithms are trained on diverse datasets and frequently examined to ensure that biases and inequalities are not perpetuated.[66]
Another future perspective on the use of AI in urology is its implementation in PCa screening programs over the next few decades. One of its main uses is to reflect on the shift to more sophisticated and personalized screening approaches using AI-driven risk-prediction models that integrate clinical data (age, family history), PSA trends, molecular analysis, genetic data, guided biopsies, and imaging (mpMRI and histopathological findings).
Future research on PCa, particularly focusing on AI-driven decision support tools, must address critical areas to ensure the accuracy, effectiveness, and broad applicability of these technologies. This involves conducting extensive validation trials to ensure that AI models can be effectively implemented in real-world settings and to assess their long-term outcomes across various clinical environments.

6. Conclusions
The diagnosis of PCa is undergoing a paradigm shift from PSA-centric approaches to a future defined by AI, novel biomarkers, and advanced imaging. These innovations promise greater precision, improved risk stratification, and reduced overtreatment; however, their widespread adoption requires multidisciplinary collaboration, regulatory oversight, and equitable implementation strategies. Artificial intelligence alone cannot substitute for existing standards of care, but can and should be positioned to act as an adjunct, enhancing its collective value and supporting clinicians in tailoring diagnostic strategies. As we move forward, embracing multimodal diagnostic models will be essential for delivering personalized data-driven care to men at risk of PCa.

Acknowledgments
The authors thank the Department of Urology at Hospital Clínico San Carlos, Madrid, for their support during the preparation of this manuscript. Special thanks go to the Universidade Federal do Rio Grande do Sul and the Hospital de Clínicas de Porto Alegre for providing academic and research resources.

Statement of ethics
Not applicable.

Conflict of interest statement
The authors declare that they have no conflicts of interest.

Funding source
None.

Author contributions
HMFSS, JGR, PMD: Conception and design;
PMD, MJM: Administrative support;
HMFSS, JGR: Provision of study materials or patients;
HMFSS, CG-S, LFM, IG, and
JMS: Collection and assembly of data;
HMFSS, JGR, PMD, MJM: Data analysis and interpretation; All authors: Manuscript writing and final approval of the manuscript.

Data availability
Data sharing is not applicable to this article as no new data were created or analyzed in this study.