Standardized MRI staging for bladder cancer relies on contrast injection. The Vesical Imaging Reporting and Data System (VI-RADS) was proposed in 2018 as a structured 1-to-5 scoring framework for evaluating multiparametric MRI findings in bladder cancer, focusing specifically on determining whether a tumor has invaded the muscular wall of the bladder, a distinction critical for selecting radical versus bladder-preserving treatment.
Conventional VI-RADS uses three imaging sequences: T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and dynamic contrast-enhanced imaging (DCEI). However, DCEI requires intravenous administration of gadolinium-based contrast agents, which carry risks including allergic reactions, nephrogenic systemic fibrosis in patients with impaired kidney function, and accumulation of gadolinium in brain structures with repeated use.
Many bladder cancer patients are elderly and have comorbid chronic kidney disease, making contrast administration risky or contraindicated. An equivalent non-contrast-enhanced VI-RADS system using only T2WI and DWI would extend MRI-based staging to patients currently excluded from it, while also reducing scan time and cost for all patients.
AI-powered noise reduction for higher-resolution MRI. Denoising deep learning reconstruction (dDLR) is a technology applied retrospectively to MRI images to improve their signal-to-noise ratio (SNR) without requiring additional scan time. A deep learning model trained on large image datasets learns to distinguish true anatomical signal from noise, enabling reconstruction of cleaner images from data already acquired during the standard scan.
Next-generation 3-Tesla MRI scanners with high-gradient magnetic fields of up to 100 mT/m can acquire thinner image slices at the same bandwidth, enabling spatial resolutions not achievable on conventional scanners. Combined with dDLR, these high-gradient (HG) scanners can produce 2-millimeter-thin slices with preserved SNR, compared to the 3-to-4-millimeter slices typically required to maintain image quality on standard scanners.
The hypothesis tested in this study was that the combination of high-gradient MRI and dDLR might provide sufficient image quality from T2WI and DWI alone to replace the information provided by DCEI, thereby enabling accurate non-contrast-enhanced VI-RADS scoring.
163 patients enrolled with full multiparametric MRI before surgery. From January 2019 through December 2020, 163 consecutive patients undergoing bladder MRI before transurethral resection of bladder tumor (TURBT) at Kyorin University School of Medicine were prospectively enrolled, with 108 patients confirmed to have urothelial bladder cancer by pathology included in the final analysis.
All scans used the same high-gradient 3-T MRI scanner (Vantage Galan 3T/ZGO, Canon Medical Systems) incorporating the Advanced intelligent Clear-IQ Engine (AiCE) dDLR technology. The scan protocol included 2-millimeter and 4-millimeter slice T2WI, 1.5-millimeter and 4-millimeter slice DWI, and 1-millimeter slice DCEI, enabling direct comparison of imaging at different resolutions and with or without deep learning reconstruction.
Two independent readers scored all images: reader 1, a board-certified radiologist with 7 years of urogenital radiology experience, and reader 2, a senior radiology resident with 1 year of urogenital experience. Pathological diagnosis from TURBT specimens served as the reference standard for muscle invasion status.
Head-to-head comparison of three VI-RADS variants. Three scoring approaches were assessed in parallel for each patient. Conventional VI-RADS incorporated all three imaging types (T2WI, DWI, and DCEI) and used a threshold of score 4 or higher to classify muscle invasion as positive, consistent with existing literature recommendations for maximizing specificity.
Non-contrast-enhanced VI-RADS (NCE-VI-RADS) used only T2WI and DWI without DCEI. It was scored as negative if both T2WI and DWI categories were below 4, and as positive if either T2WI or DWI (or both) scored 4 or above. This asymmetric combination rule was designed to preserve sensitivity for muscle invasion when the DCEI confirmation step was absent.
A third variant, NCE-VI-RADS with dDLR, replaced standard 4-millimeter T2WI with 2-millimeter T2WI processed through deep learning reconstruction, while retaining standard DWI. This exploratory variant tested whether the improved spatial resolution and noise reduction from dDLR could incrementally improve NCE-VI-RADS performance.
NCE-VI-RADS matched conventional VI-RADS accuracy. Muscle invasion was confirmed pathologically in 23 of 108 patients (21%). AUC values for diagnosing muscle invasion were 0.94 and 0.91 for conventional VI-RADS (readers 1 and 2), and 0.93 and 0.91 for NCE-VI-RADS (readers 1 and 2), with no statistically significant difference between the two systems for either reader (reader 1: p = 0.08; reader 2: p = 0.95).
NCE-VI-RADS with dDLR achieved AUC values of 0.96 and 0.93 for readers 1 and 2 respectively, numerically exceeding conventional VI-RADS for both readers, though these differences also did not reach statistical significance (reader 1: p = 0.48; reader 2: p = 0.54). These results suggest that the addition of deep learning reconstruction may provide further diagnostic improvement.
T2WI alone achieved AUC of 0.88 to 0.89, DWI alone achieved 0.90 to 0.94, and DCEI alone achieved 0.94 to 0.96 for individual readers. The combined NCE-VI-RADS score using DWI and T2WI together reached the same level as DCEI-inclusive scoring, indicating that the combination of T2WI and DWI provides sufficient diagnostic information to replace DCEI in most cases.
Deep learning reconstruction improved reader consistency. Inter-reader agreement for NCE-VI-RADS showed substantial agreement by kappa statistics (kappa 0.61 to 0.80), compared to almost perfect agreement (kappa 0.81 to 1.00) for conventional VI-RADS. This gap suggests that without DCEI, experienced and less experienced readers differ somewhat more in their interpretation of ambiguous cases.
Importantly, NCE-VI-RADS with dDLR achieved almost perfect inter-reader agreement at the same level as conventional VI-RADS, indicating that the improved image clarity from deep learning reconstruction reduced ambiguity for the less experienced reader, effectively closing the agreement gap that existed with standard NCE-VI-RADS.
The improvement in inter-reader agreement with dDLR is clinically significant because it suggests that deep learning reconstruction could make non-contrast VI-RADS more reproducible across radiologists with different levels of urogenital MRI experience, supporting broader adoption in centers without subspecialty-trained readers.
First prospective validation of a contrast-free VI-RADS system in a large cohort. This study was the largest prospective validation of biparametric (non-contrast) MRI for VI-RADS scoring in bladder cancer, enrolling three times as many participants as the only prior study examining contrast-free VI-RADS assessment. The results confirm that DCEI can be safely omitted from VI-RADS assessment when both T2WI and DWI are included.
For patients with contraindications to gadolinium-based contrast agents, including those with impaired renal function or prior allergic reactions, NCE-VI-RADS provides a clinically valid alternative that maintains diagnostic accuracy. Broader adoption of NCE-VI-RADS would also reduce scan duration and eliminate contrast agent costs for all patients undergoing bladder MRI staging.
Key limitations include the single-institution design, the relatively small sample size of 108 patients, and the use of a high-gradient MRI scanner not yet widely available. The dDLR findings are exploratory given the lack of statistical significance and require confirmation in multi-institutional prospective studies with larger populations before clinical recommendation.