Population, inclusion, and exclusion criteria
This study was compliant with Helsinki Declaration. The following inclusion and exclusion criteria were considered: Inclusion criteria: patients who underwent 3-T multiparametric MRI (mpMRI) at our institute, including two sets of repeated DWI acquisitions for evaluating prostate lesions with informed consent from July 2016 to May 2020. Exclusion criteria: treatment except radical prostatectomy; lesions with a longitudinal diameter < 10 mm; lesions not detected on DWI; lesions with a voxel number within the region of interest (ROI) < 50; lesions containing voxel with ADC value < 0; poor image quality. Figure 1 shows the flowchart of patient inclusion and exclusion.
MRI
MRI was performed using a 3-T system (Ingenia, Philips Healthcar, Eindhoven, The Netherlands) with a pelvic phased-array coil. No endorectal coil was used. Either 20-mg hyoscine-N-butyl-bromide or 1-mg glucagon was injected intramuscularly before examination to minimize bowel peristalsis.
A routine mpMRI protocol was applied to all patients, including sagittal, coronal, and axial T2-weighted imaging; axial DWIs; and axial dynamic contrast-enhanced imaging before and after gadolinium chelate injection of 0.1 mmol/kg gadoterate meglumine, Magnescope, Dotarem (Guerbet, Villepinte, France). For DWI, two sequential free-breathing DWI single-shot spin-echo echo-planar images were acquired. The patient remained in the same position between the two DWI acquisitions. Four b-values (0, 100, 1,000, and 1,500 s/mm2) with three orthogonal diffusion probing gradients were generated. ADC maps were generated using DWIs with b-values of 100 and 1,000 s/mm2, ADC map (100, 1,000) in line with the Prostate Imaging–Reporting and Data System (PI-RADS) version 2.1 (https://www.acr.org/-/media/ACR/Files/RADS/PI-RADS/PIRADS-V2-1.pdf) for the first and second DWIs, respectively. The DWI sequence parameters are summarized in Supplemental Table S1.
Image analysis
Image analysis including ROI assignment was performed by a consensus decision of two observers (C.T. and M.H. with 4 and over 30 years of experience in diagnostic radiology, respectively) using a Synapse Vincent 3D Image Analysis System (Fujifilm Corporation, Tokyo, Japan). For PZ cancer, the polygonal two-dimensional ROI was manually determined on the lesion in the center slice showing hyperintensity on the first DWI with a b-value of 1,500 s/mm2 (DWI 1,500) and hypointensity on the first ADC map (100, 1,000), referring to T2-weighted imaging, dynamic contrast-enhanced imaging, and whole-mount, step-sectioned histological evaluation of prostatectomy specimen. Then, the ROI was placed on the first DWI datasets of DWI 0, DWI 100, and DWI 1,000 through IVIM application of a Synapse Vincent 3D Image Analysis System. For non-peripheral transition zone (TZ) cancers, the polygonal two-dimensional ROI was manually determined on the lesion in the center slice showing hypointensity on T2-weighted images and hyperintensity on the first DWI 1,500, referring to the first ADC map (100, 1,000), dynamic contrast-enhanced imaging, and whole-mount, step-sectioned histological evaluation of prostatectomy specimen. After this, the ROI was placed on the first DWI datasets of DWI 0, DWI 100, and DWI 1,000 through intravoxel incoherent motion (IVIM) application of the Synapse Vincent 3D Image Analysis System. The same procedures were repeated for the second DWI datasets. Voxel data distributions within the ROI were rendered in comma-separated values (CSV) format (Supplemental Figs. S1 and S2) using a Synapse Vincent 3D Image Analysis System. Then, the ADC of each voxel was calculated by fitting signal intensity decay between four patterns of b-value combinations using a monoexponential curve fit: 0 and 1,000 s/mm2, ADC (0, 1,000); 0 and 1,500 s/mm2, ADC (0, 1,500); 100 and 1,000 s/mm2, ADC (100, 1,000); and 100 and 1,500 s/mm2, ADC (100, 1,500). Representative cases are shown in Figs. 2 and 3.
We assigned a two-dimensional ROI in the center slice of the lesion because more than half lesions were not large enough to place a three-dimensional ROI. Only 21 lesions, 43% of total lesions, showed longitudinal diameter > 12 mm on images and could be determined on equal to or more than four slices (DWI slice thickness of 3 mm/ gap of 0 mm, Supplemental Table 1) that would have satisfied assigning ROIs on two or more slices avoiding peripheral images, possibly being affected by partial volume effect. Texture analysis calculates the relationship between adjacent voxels, and thus, we assumed that appropriate texture analysis required at least four voxels along each direction.
All voxels within the ROI were extracted from the CSV data of ADC (0, 1,000), (0, 1,500), (100, 1,000), and (100, 1,500), and DWI 0, 100, 1,000, and 1,500. First-order statistical variables (minimum, 10%, 25%, median, 75%, 90%, maximum, mean, sum, standard deviation, skewness, kurtosis, energy, and entropy) were calculated. After discretization of voxel values (bin number 32), higher-order texture analysis was performed in a two-dimensional manner to generate a gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level zone-size matrix (GLZSM), and neighborhood gray-level difference matrix (NGLDM). Homogeneity, energy, correlation, contrast, entropy, and dissimilarity were calculated from the GLCM. Short-run emphasis (SRE), long-run emphasis (LRE), low gray-level run emphasis (LGRE), high gray-level run emphasis (HGRE), short-run low gray-level emphasis (SRLGE), short-run high gray-level emphasis (SRHGE), long-run low gray-level emphasis (LRLGE), long-run high gray-level emphasis (LRHGE), gray-level non-uniformity for run (GLNUr), run-length non-uniformity (RLNU), and run percentage (RP) were calculated from GLRLM. Short-zone emphasis (SZE), long-zone emphasis (LZE), low gray-level zone emphasis (LGZE), high gray-level zone emphasis (HGZE), short-zone low gray-level emphasis (SZLGE), short-zone high gray-level emphasis (SZHGE), long-zone low gray-level emphasis (LZLGE), long-zone high gray-level emphasis (LZHGE), gray-level non-uniformity for zone (GLNUz), zone length non-uniformity (ZLNU), and zone percentage (ZP) were calculated from GLZSM. Coarseness, contrast, and busyness were calculated from NGLDM. Texture features were computed using the PTexture package (www.github.com/metavol/ptexture) written in Python language. The detailed methods are described elsewhere [16].
Statistical analysis
Statistical analyses were carried out separately for PZ and TZ cancers. First, the correlation between texture features and GG was evaluated using Spearman's rank correlation test. For the features showing significance at both the first and second examinations, receiver operating characteristic (ROC) curves for differentiating between GG of 1 and 2 versus GG of 3, 4, and 5 were drawn, and the area under the curve (AUC) was calculated because there was a difference in prognosis between GG of 1 and 2, versus GG of 3, 4, and 5 [17]. To check test-retest data repeatability, intraclass correlation coefficient (ICC) and Bland-Altman plot (%) (% difference was used to normalize differences in original data magnitude) were used. Statistical analyses were performed using GraphPad Prism ver. 7.05 (GraphPad Software, San Diego, USA) and SPSS statistics ver. 25 (International Business Machines Corporation, Armonk, USA); p-values < 0.05 were considered statistically significant.
We considered the following values for classifying the strength of correlation: moderate (|ρ|: 0.4-0.7), strong (|ρ|: 0.7-0.9), and very strong (|ρ|: 0.9-1) [18], discrimination ability: acceptable (AUC: 0.7–0.8), excellent (AUC: 0.8–0.9), and outstanding (AUC > 0.9) [18], and data repeatability: moderate (ICC: 0.5–0.75), good: (ICC: 0.75–0.9), and excellent (ICC: 0.9–1) [19].