Prostate cancer is the second most common cancer in men worldwide and a leading cause of cancer-related death. Despite its prevalence, no widely accepted screening program exists, making high-quality diagnostic imaging critically important for detection, staging, and treatment planning.
Multiparametric MRI (mpMRI) has become the standard diagnostic tool, guided by the Prostate Imaging Reporting and Data System (PI-RADS). A standard mpMRI protocol takes 30 to 45 minutes to complete, with T2-weighted imaging (T2WI) alone accounting for roughly 60% of that time.
Long scan times create multiple problems: patients experience discomfort and anxiety, which can cause motion that degrades image quality. From a system perspective, long scans limit the number of patients who can be scanned each day, contributing to lengthy waiting lists and higher costs.
Previous technologies such as parallel imaging and compressed sensing attempted to reduce scan times, but both had significant limitations. Compressed sensing in particular is computationally demanding, generates its own artifacts, and has proven clinically impractical in many settings due to long reconstruction times.
Artificial intelligence, and specifically deep learning, offers a fundamentally different approach to speeding up MRI. Rather than simply collecting fewer data points (which degrades quality), AI algorithms can learn to intelligently fill in missing information and reconstruct high-quality images from undersampled raw data.
In MRI, raw data is collected in a mathematical domain called k-space. AI models trained on thousands of image pairs -- undersampled and fully sampled -- can learn the patterns needed to reconstruct diagnostically useful images from data collected much faster. This is conceptually different from older acceleration methods that rely on mathematical assumptions rather than learned patterns.
Beyond acquisition speed, AI holds promise for improving reconstruction quality, reducing noise and artifacts, and potentially assisting with interpretation tasks such as distinguishing benign from malignant tissue and predicting tumor aggressiveness -- capabilities that could reduce the variation in diagnosis between different radiologists.
This systematic review was conducted to evaluate the current evidence for AI-accelerated prostate MRI, focusing specifically on three questions: How much does AI reduce acquisition time? Does AI affect qualitative image quality characteristics? And does AI change measurable signal properties of the images?
The review was conducted following PRISMA 2020 guidelines (Preferred Reporting Items for Systematic Reviews and Meta-Analyses), the international standard for transparent and reproducible systematic review reporting. Study quality was additionally assessed using the CLAIM checklist (Checklist for Artificial Intelligence in Medical Imaging).
Searches were conducted across three major medical databases -- PubMed, Embase, and Scopus -- covering publications from 2013 through August 2023. Search terms included combinations of artificial intelligence, machine learning, deep learning, MRI, prostate, and acceleration-related terms.
From 152 initially identified papers, 65 duplicates were removed. Of the 87 screened abstracts, 75 were excluded, and a further 4 were removed after full-text review due to incorrect interventions such as combined AI and compressed sensing approaches. Eight studies met all inclusion criteria and were included in the final review.
Inclusion required that studies used AI specifically for MRI acceleration (not just interpretation), with primary outcomes of acquisition time and image quality. Excluded were case reports, editorials, reviews, and studies of non-prostate MRI. Due to the substantial heterogeneity across studies, a narrative synthesis was chosen over formal statistical pooling.
The eight included studies spanned multiple countries (Germany, USA, South Korea, UK, and Taiwan) and were published between 2021 and 2023. Participant numbers ranged from 30 to 109, with patient ages typical for prostate cancer screening populations (mean ages 65-72 years). All studies used 3T MRI scanners, with Siemens MAGNETOM being the most common platform.
All studies incorporated T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI), the two most important sequences for prostate lesion detection. Some studies evaluated AI acceleration in T2WI only, some in DWI only, and one in both. No fewer than two radiologist readers evaluated images in any study, with reader experience ranging from 1 to 14 years.
A variety of AI architectures were employed across studies: convolutional neural networks (CNN), variational neural networks (VNN), unrolled variational networks (UVN), and commercially available deep learning reconstruction tools from scanner manufacturers such as GE Healthcare's AIR Recon DL and a Siemens commercial prototype.
This diversity in AI models, scanner types, and imaging protocols reflects the rapidly evolving nature of the field. While it demonstrates broad interest and development activity, it also complicates direct comparison between studies, as each algorithm has different training approaches, optimization goals, and architectural properties.
The most consistent finding across all eight studies was a meaningful reduction in MRI acquisition time. Reported reductions ranged from approximately 33% to over 80% compared with conventional protocols. These are not marginal improvements -- an 80% reduction means a scan that previously took 20 minutes taking just 4 minutes.
Specific examples: Johnson et al achieved an 81.5% reduction in axial imaging time. Kim et al demonstrated reductions of 68.9% for high-resolution and 75.5% for low-resolution imaging. Gassenmaier et al reported reductions of around 64% in axial imaging. Even the more modest studies achieved reductions in the 32-38% range.
Critically, all studies found these time reductions were achievable without compromising overall image quality. This combination -- faster acquisition with maintained or improved quality -- is the key promise of AI-accelerated MRI and represents a genuinely important advance over previous acceleration technologies.
From a healthcare systems perspective, these time savings translate directly into the ability to scan more patients per scanner per day, reducing wait times and potentially improving access to prostate MRI screening and diagnosis. The economic implications of faster, cheaper scans could be substantial.
Beyond overall quality, studies measured specific image quality characteristics including artifact levels, noise, image sharpness, and signal-to-noise ratio (SNR). Results here were less consistent than for acquisition time.
For artifacts in T2WI, the two Gassenmaier studies reported statistically significant improvements with AI acceleration. Kim et al and Lee et al found mostly no significant difference. For DWI, no study found any significant change in artifact levels, suggesting the effect may depend on the specific algorithm and imaging context.
For SNR (signal-to-noise ratio -- a measure of how clearly the tissue signal stands out from background noise) results were notably mixed. Park et al reported significantly higher SNR with AI. Kim et al and Lee et al both reported significantly lower SNR when using AI acceleration alone, but Lee et al crucially showed that applying AI denoising as a post-processing step significantly increased SNR above conventional levels. This suggests AI can work at two stages: during acquisition for speed, and afterward for quality enhancement.
Results for contrast-to-noise ratio (CNR) -- the ability to differentiate adjacent tissues -- were similarly mixed, with some studies reporting improvement and others showing reduction. Several studies noted significant variation between different radiologist readers in their quality assessments, underscoring the subjective component in radiological image evaluation.
The review authors identify several factors that likely explain the variable results across studies. Different AI architectures have different strengths -- a network optimized for noise reduction may introduce different trade-offs than one optimized for artifact suppression. The training data used to build each model also shapes its behavior, and models trained on narrow or homogeneous datasets may perform well in their original context but fail to generalize.
Reader variability adds another layer of complexity. With experienced radiologists interpreting images differently than less experienced ones, and with image quality being partly subjective, the same AI-accelerated image can receive different quality scores from different readers. This makes cross-study comparison challenging.
The finding that AI denoising can recover or even improve SNR and CNR in DWI sequences is particularly noteworthy. It suggests that AI-accelerated MRI workflows might benefit from a two-step approach: rapid AI-assisted acquisition followed by AI-based reconstruction enhancement, potentially delivering both speed and quality improvements simultaneously.
Standardization emerges as a key need. Without agreed metrics, protocols, and reporting standards, accumulating comparable evidence across studies is difficult. The field would benefit from consensus on what outcomes to measure, how to measure them, and how to evaluate AI models across different clinical contexts and scanner platforms.
The systematic review concludes that AI-acceleration in prostate MRI shows genuine promise, with the universal finding of meaningful scan time reductions without compromising overall image quality representing a clinically significant advance. All eight included studies met quality standards and demonstrated low risk of bias.
The potential clinical benefits are broad: reduced scan times mean less patient discomfort and anxiety, fewer motion artifacts, more patients scanned per day, shorter waiting lists, and potential cost reductions. For healthcare systems facing growing demand for prostate MRI, these efficiency gains could have major practical impact.
The authors note a particularly exciting long-term possibility: if AI-accelerated MRI can be validated in large-scale studies, it could support the development of a prostate cancer screening program. Currently, no acceptable mass screening modality exists for prostate cancer. A faster, cheaper, high-quality MRI protocol could fill this critical gap in early detection.
Before wide clinical adoption, however, larger and more diverse studies are needed. The current evidence base is limited to eight studies with relatively small sample sizes. Standardized evaluation frameworks, diverse training datasets, and long-term follow-up data will all be required before these tools can be confidently recommended for routine clinical deployment.