Participants with suspected bladder cancer who underwent a 3-T bladder MRI between August 2022 and February 2023 were consecutively enrolled. This observational prospective single-center study obtained Ethical approval. The study was conducted in line with the Declaration of Helsinki and its subsequent revisions, and written informed consent was obtained from all participants. Inclusion criteria were as follows: (1) Bladder tumor identified for the first time, with no prior treatment; (2) No bladder biopsy conducted within 2 weeks before MRI assessment; (3) Absence of contraindications for MRI examination. Exclusion criteria: (1) Patients who did not undergo surgery intervention; (2) Pathological confirmation of non-urothelial bladder cancer. The participant selection process is shown in Fig. 1. The data and material for this study are not available due to possible compromise of personal privacy.

Flowchart shows the number of participants recruited and number and reason for exclusion from study

All MRI examinations were performed on a 3-T MRI system (SIGNA Architect, GE Healthcare) with an AIR anterior array coil. Participants were instructed to void their bladders two hours before the imaging. For patients experiencing frequent urination, a water intake of 500–1000 mL was advised 30 min before the examination. Those without contraindications for spasmolytic treatment received a 1 mL intramuscular injection of scopolamine butylbromide.

The multiparameter MRI protocol included the following sequences: axial, coronal, and sagittal T2-weighted imaging (T2WI) sequence, axial fFOV DWI, standard rFOV (rFOVSTA) DWI followed by fast rFOV with DLR (rFOVDLR) DWI with similar acquisition parameters and reduced numbers of excitation, axial dynamic contrast-enhanced imaging (DCEI). The b-values were 50 s/mm2 and 1000 s/mm2 for three DWIs. Apparent diffusion coefficient (ADC) maps were calculated for each DWI. The scan time of fFOV DWI, rFOVSTA DWI, and rFOVDLR DWI were 1:39, 5:02, and 3:25 min, respectively. FOCUS DWI was performed as rFOV DWI in our study. Detailed image parameters and time are displayed in Table 1.

The AIRTM Recon DL algorithm (GE Healthcare) based on feedforward deep convolutional neural networks was used to reconstruct rFOVDLR DWI. Convolutional neural networks accept raw unfiltered complex-valued input data and provide output images with improved signal-to-noise ratio [23]. The software provides a user-specified denoising level from 0% to 100%, where 0% means conventional reconstruction without DL; other options are as follows: low (33%), medium (50%), and high (75%). In the present study, a 75% noise reduction factor was chosen. The detailed network design and performance in phantom images are shown in the white paper [23].

Two genitourinary radiologists (reader 1 Y.C., and reader 2 X.X.Z., with 29, and 4 years of experience in abdominal MRI, respectively) independently reviewed fFOV DWI, rFOVSTA DWI, and rFOV DLR DWI in random order during separate sessions, with a month interval between sessions. All image analyses were performed on AW 4.7 workstation (GE Medical Systems). The presenter of the images (Y.C.W., with 9 years of experience in abdominal MR) recorded the reader’s rating results of imaging quality assessment and VI-RADS scoring. In cases with multiple lesions, the lesion with the greatest invasion depth or largest size (in cases of equal degrees of invasion) was selected by a radiologist (X.X.J., with 18 years of experience in abdominal MR) before assessment.

Qualitative evaluation was performed using a 4-point scoring system. The evaluation criteria are as follows: overall image quality (1 = poor image quality; 2 = fair image quality; 3 = good image quality; 4 = excellent image quality), motion artifacts (1 = severe artifact with no diagnostic value; 2 = moderate artifact with effect on diagnostic assessment; 3 = mild artifact without interference of diagnostic assessment; 4 = no artifact), bladder wall sharpness (1 = severe blurring, 2 = intermediate blurring, 3 = slight blurring, 4 = no blurring).

For quantitative evaluation, oval regions of interest were manually drawn on the iliopsoas muscle and the lesion in a single representative slice of DWIs (b = 1000 s/mm2), and automatically copied to the ADC maps. The SNR of the lesion and the contrast-to-noise ratio (CNR) of the lesion to the iliopsoas muscle were calculated according to the following equations:

$$=\frac}_}-}_})}}_}}^+}_}}^)}}$$

SItumor and SDtumor represent the mean and standard deviation values of signal intensity of the tumors respectively, while SImuscle and SDmuscle represent the mean value and standard deviation of signal intensity of the iliopsoas muscle, respectively.

All MRI images were independently evaluated according to VI-RADS [3] by the above two readers without knowledge of the surgical or histologic findings. Category by single sequence (T2WI, fFOV DWI, rFOVSTA DWI, rFOV DLR DWI, and DCEI) was separately assessed with an interval of two weeks between each sequence. And the final VI-RADS score of set1, set2, and set3 was assigned. Each set included axial, coronal, and sagittal T2W images, DCEI, and DWI with the corresponding ADC map. In detail, fFOV DWI was included in set 1, rFOVSTA DWI in set 2, and rFOV DLR DWI in set 3. The 5-point scores using VI-RADS were compared with the pathological results of surgery.

The ADC values were measured by using a single representative slice of the tumor. Regions of interest were manually drawn on fFOV DWI, and were copied to rFOVSTA DWI, and rFOVDLR DWI with a b value of 1000 s/mm2, and the mean ADC of the ROI was recorded. Tumor stalk or thickened submucosa and vessels were excluded using T2WI as a reference.

All patients underwent transurethral resection of bladder tumor or radical cystectomy within four weeks after MRI. When patients had both, radical cystectomy was considered as the final standard of reference. According to European association of Urology guidelines, a second TURB may be performed for high-risk patients [24].

The histological type, grade, and stage of the tumors were assessed by pathologists according to the 2016 World Health Organization grading systems and the 2017 American Joint Committee on Cancer/Union for International Cancer Control TNM staging system.

The sample size of this study was calculated by comparing SNR and CNR means between fast sequence with DLR and standard sequence. A confidence interval of 95% and a power of 90% was considered. Details information on the sample size calculation and the tool used can be found in supplement S1 and Table S1 (supplement online). The number of patients needed in this study to obtain the desired power was 68.

The Kolmogorov–Smirnov test was used to test the normal distribution of quantitative data and Likert scales. This test showed that the distribution of the values of SNR, CNR, ADC value, and Likert scales of image quality were non-normal. Therefore, quantitative data and Likert scales were compared by Friedman test with Dunn’s pairwise post hoc test. Bonferroni correction p values for multiple comparisons were applied. Receiver operating characteristic curve analysis was used to analyze the accuracies of VI-RADS in predicting muscle invasion. The optimal cutoff value of the VI-RADS score was determined by maximization of Youden’s index. Sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and area under the curve (AUC) were calculated for all radiologists. Delong’s test was used to calculate the difference between every two groups of AUC. Intraclass correlation coefficients were used to evaluate interobserver agreement for SNR, CNR, and ADC value. Additionally, interobserver agreements for qualitatively assessed image quality and VI-RADS score were evaluated through Cohen κ. The κ values were interpreted as follows: 0–0.20 = poor agreement, 0.21–0.40 = fair agreement, 0.41–0.60 = moderate agreement, 0.61–0.80 = substantial agreement, and 0.81–1 = excellent agreement. All statistical analyses were performed using the software SPSS version 27.0 (IBM). All tests were two-sided and statistical significance was determined to be p < 0.05.