Long MRI acquisition times continue to challenge prostate imaging, particularly when T2-weighted sequences remain central to transition-zone lesion assessment and local staging. AI reconstruction offers a route to shorter scans, but its diagnostic performance under clinical conditions remains an important question. A recent analysis published in European Radiology addresses that question in accelerated prostate MRI using clinically acquired multi-coil imaging rather than simulated or single-coil data alone. The comparison covers conventional imaging alongside threefold and sixfold acceleration, with interpretation by eight experienced prostate radiologists. It also examines performance within a biparametric protocol that includes diffusion-weighted imaging and apparent diffusion coefficient maps, using histopathology and PI-RADS findings as reference standards.
Clinical Dataset and Reconstruction Approach
The testing cohort comes from consecutive biparametric prostate MRI examinations acquired in routine clinical practice between May 2022 and January 2024. After exclusions, 120 baseline cases remain for evaluation. Each examination includes axial T2-weighted imaging, diffusion-weighted imaging and an apparent diffusion coefficient map. Positive cases are defined as Gleason grade group 2 to 5, while negative cases are based on PI-RADS 2 or lower and or biopsy-negative findings. The non-accelerated T2-weighted scan acts as the reference condition for both diagnostic assessment and image-quality review.
Must read: Radiomics AI Supports Detection of Prostate Cancer
Acceleration is created retrospectively from the clinical data. The reference scan is treated as R = 1, with additional reconstructions at R = 3 and R = 6. The threefold setting uses one of three signal averages. The sixfold setting combines a single average with a higher parallel imaging factor. That shift reduces estimated T2-weighted acquisition time from several minutes in the reference condition to much shorter time ranges in the accelerated settings, reaching well below two minutes at the highest level and as low as under one minute in some cases. Diffusion-weighted imaging is not undersampled.
The reconstruction method is vSHARP, retrained specifically on T2-weighted prostate MRI data. The model combines deep learning priors with physics-based data consistency in an iterative framework. The work therefore centres on clinically feasible scan acceleration using routine acquisition modifications, followed by reconstruction designed to preserve anatomical reliability rather than simply maximise visual similarity.
Reader Design and Diagnostic Performance
Diagnostic performance is assessed in a partially paired multi-reader multi-case design. The eight participating radiologists all have extensive prostate MRI experience, and each has interpreted more than 1000 clinical prostate MRI cases. To distribute workload, the dataset and readers are divided into three groups. Readings take place across three sessions spaced four weeks apart to reduce recall bias. Within each session, radiologists first assess the T2-weighted image for visual quality and diagnostic metrics, then immediately repeat diagnostic assessment for the same case with the standard diffusion-weighted and apparent diffusion coefficient images, without revisiting the initial T2-weighted judgement.
Across the T2-weighted protocol, diagnostic performance does not differ significantly between the acceleration conditions. AUROC values decline from 0.86 at R = 1 to 0.82 at R = 3 and 0.80 at R = 6, with a comparison across conditions yielding p = 0.08. The direction of change points downward, but the reduction does not reach statistical significance. For the full biparametric protocol, performance remains slightly higher and more stable, with AUROC values of 0.88, 0.86 and 0.85 and no significant difference across conditions.
PI-RADS-based sensitivity also remains high. In the T2-weighted protocol, sensitivity stays above 0.90 across all three conditions, while specificity changes only modestly. In the biparametric setting, sensitivity remains 0.97 across acceleration levels, with specificity again varying within a narrow range. Inter-reader agreement falls at higher acceleration for T2-weighted imaging, while the combined biparametric protocol shows greater stability. The pattern suggests preserved discrimination overall, alongside some loss of consensus as acceleration increases.
Image Quality, Practical Gains and Study Limits
Perceived image quality is not reduced by moderate acceleration and, in some respects, improves. At R = 3, radiologists rate sharpness and noise significantly better than in the non-accelerated reference. Artefacts, lesion conspicuity and overall visual quality also move in a favourable direction at that setting. At R = 6, perceived quality remains statistically comparable with the reference condition across the rated dimensions. The most pronounced gains therefore appear at the intermediate acceleration level rather than the highest one.
Objective image-quality measures follow the same broad pattern. Structural similarity remains high at both accelerated settings, while signal-to-noise ratio declines only modestly at the highest acceleration. Error measures rise slightly at R = 6, indicating minor reconstruction discrepancies and some loss of fine detail. Those shifts align with the reader-based assessment, which shows preserved quality overall but less advantage at the strongest acceleration.
The discussion links shorter scan times with shorter examinations, lower risk of motion artefacts and more efficient use of MRI capacity. At the same time, several constraints remain clear. All testing data come from a single vendor platform. Acceleration is simulated retrospectively rather than acquired prospectively. The reading environment differs from routine hospital workflow, and the reference reconstruction omits vendor-specific preprocessing to keep conditions uniform. Negative cases do not include long-term follow-up, image-quality assessment is limited to T2-weighted sequences, and the cohort contains relatively few transition-zone lesions. Reader-to-reader variability, especially for specificity, also remains substantial.
AI reconstruction enables marked acceleration of T2-weighted prostate MRI while maintaining diagnostic performance close to the non-accelerated reference in this multi-reader evaluation. The strongest visual benefit appears at threefold acceleration, while sixfold acceleration preserves perceived quality and keeps cancer detection broadly comparable without a statistically significant drop. The results still show a downward movement in diagnostic performance and reader agreement as acceleration increases. That pattern supports further prospective evaluation while indicating that accelerated reconstruction can shorten prostate MRI substantially within the tested range.
Source: European Radiology
Image Credit: iStock
References:
van Lohuizen Q, Fransen SJ, Yiasemis G et al. (2026) Diagnostic assessment of artificial intelligence reconstruction on accelerated prostate MRI: a retrospective, paired, multi-reader multi-case study. Eur Radiol: In Press.
