A Narrative Review of Artificial Intelligence in MRI-Guided Prostate Cancer Diagnosis: Addressing Key Challenges

Abstract

Background/Objectives: Magnetic resonance imaging (MRI) is crucial in detecting suspicious lesions and diagnosing clinically significant prostate cancer (csPCa). However, variability in MRI-targeted diagnostic pathways arises due to factors such as patient characteristics, imaging protocols, and radiologist expertise. Artificial intelligence (AI) offers potential solutions to these challenges by enhancing diagnostic accuracy and efficiency. Methods: This narrative review explores AI techniques, particularly machine learning and deep learning, in the context of prostate cancer diagnosis. It examines their application in improving MRI scan quality, detecting artifacts, and assisting radiologists in lesion detection and interpretation. It also considers how AI helps to reduce reading time and inter-reader variability. Results: AI has demonstrated sensitivity that is generally comparable to experienced radiologists, although specificity tends to be lower, potentially increasing false-positive rates. The clinical impact of these results requires validation in larger, prospective multicenter studies. AI is effective in identifying artifacts, assessing MRI quality, and assisting in diagnostic efficiency by providing second opinions and automating lesion detection. However, variability in study methodologies, datasets, and imaging protocols can impact AI's generalizability, limiting its broader clinical application. Conclusions: While AI shows significant promise in enhancing diagnostic accuracy and efficiency for csPCa detection, challenges remain, particularly with the generalizability of AI models. To improve AI robustness and integration into clinical practice, multicenter datasets and transparent reporting are essential. Further development, validation, and standardization are required for AI's widespread clinical adoption.

Keywords: artificial intelligence (AI); deep learning (DL); machine learning (ML); magnetic resonance imaging (MRI) of prostate; prostate cancer (PCa).

Figure 1

Despite advancements in multi-parametric magnetic resonance imaging (mpMRI) technology, significant variability persists in magnetic resonance imaging (MRI) acquisition, interpretation, prostate imaging and reporting system (PI-RADS) scoring, and reporting, which can affect diagnostic accuracy. The MRI-first approach, recommended before biopsy in men with suspected clinically significant prostate cancer (csPCa), aims to improve early detection and reduce unnecessary biopsies. Artificial intelligence (AI) holds considerable potential to address this variability and contributes to the standardization of critical aspects such as MRI acquisition, interpretation, and reporting. Within the overall diagnostic pathway for prostate cancer (PCa) detection, orange-colored boxes represent the role of MRI guidance.

Figure 2

AI-based methods have demonstrated promising potential to enhance multiple stages of the MRI-guided prostate cancer PCa diagnostic pathway. The figure illustrates AI′s role across three main components: (1) MRI Acquisition, where AI contributes through quality assurance, artifact detection, and deep-learning-based image reconstruction to improve image clarity and reduce acquisition times; (2) MRI Interpretation, in which AI provides automated lesion detection, segmentation, classification, and serves as a second-opinion tool to reduce inter-reader variability; and (3) Reporting, highlighting AI’s capacity to standardize results, provide structured lesion quantification, and facilitate clear communication of imaging findings. Collectively, these applications underscore AI’s potential to standardize and optimize the diagnostic process in clinical practice. Orange-colored boxes indicate the key stages of prostate MRI processing—namely acquisition, interpretation, and reporting. The green boxes branching from them reflect clinical challenges associated with each stage, along with the contributions and potential solutions provided by AI.

Figure 3

T2WI (a) ADC map (b) DWI) (c) and calculated b1500 DWI (d) images of a 63-year-old man with a serum prostate-specific antigen (PSA) level of 3.7 ng/mL. A focal hypointense prostate lesion is identified in the right peripheral zone (PZ), measuring 28 mm in diameter. The lesion is markedly hypointense on the ADC map and slightly hyperintense on high b-value images. According to PI-RADS v2.1 criteria, the lesion is classified as PI-RADS 3. Additionally, another PI-RADS 3 lesion is observed in the left lateral PZ, measuring 14 mm in diameter. This lesion appears focal hypointense on T2WI, markedly hypointense on the ADC map, and is indistinguishable on high b-value diffusion-weighted images.

Figure 4

T2WI images (a) and a prostate sector map (b) showing marked prostate lesions in a 63-year-old man with a serum PSA level of 3.7 ng/mL (same patient as in Figure 3). The deep learning (DL)-based application has segmented prostate lesions in the right and left lateral PZs, assigned PI-RADS scores, and determined the dimensions and volume of the lesions. The lesion locations are also marked on the prostate sector map for potential guidance in MRI-targeted biopsy. This lesion was clinically diagnosed as focal prostatitis, and the patient was scheduled for follow-up. In this case, AI provided the necessary prostate lesion assignments in advance, potentially reducing reading time and assisting less-experienced readers in detecting lesions and assigning PI-RADS scores. This is particularly valuable because DL models are trained on expert-annotated datasets, effectively transferring the knowledge of expert prostate imaging readers to less-experienced practitioners. By doing so, AI contributes to democratizing diagnostic imaging in a scalable manner.

Figure 5

T2WI (a), ADC map (b), DWI (c), and calculated b1500 DWI (d) images of a 65-year-old man with a serum PSA level of 6.4 ng/mL. A focal hypointense prostate lesion measuring 18 mm in diameter is identified in the right PZ, appearing markedly hypointense on the ADC map and markedly hyperintense on high b-value images. Additionally, another prostate lesion measuring 17 mm in diameter is observed in the left posterolateral PZ, appearing focally hypointense on T2WI, markedly hypointense on the ADC map, and markedly hyperintense on high b-value diffusion-weighted images. According to PI-RADS v2.1 criteria, both prostate lesions are classified as PI-RADS 5. Targeted biopsy confirmed csPCa. AI-assisted prostate MRI interpretation can play a crucial role in detecting csPCa, improving risk stratification, and guiding biopsy decisions. In this case, AI could have assisted in the following: enhancing lesion detection and segmentation, ensuring both lesions were identified and properly characterized; standardizing PI-RADS assessment, reducing inter-reader variability and increasing diagnostic confidence; supporting MRI-targeted biopsy planning by accurately marking lesion locations on a prostate sector map, improving sampling precision; optimizing workflow efficiency by reducing reading time and providing quantitative lesion metrics, which may help in csPCa risk assessment and treatment planning. By leveraging AI in prostate MRI evaluation, radiologists can enhance diagnostic accuracy, minimize missed csPCa cases, and support less-experienced readers in high-stakes clinical decision-making.

Figure 6

T2WI images (a) and prostate sector map (b) showing marked prostate lesions in a 65-year-old man with a serum PSA level of 6.4 ng/mL (same patient as in Figure 5). The DL-based application has segmented prostate lesions in the right lateral PZ and left posterolateral PZ, assigned PI-RADS scores, and determined the dimensions and volume of the prostate lesions. The lesion locations are also marked on the prostate sector map for guidance in MR-targeted biopsy.