An Institutional Review Board (IRB) approved a protocol for retrospective review of mpMRI exams from prostate cancer patients. The IRB waived the need for informed consent. A total of 71 sequential patients, presenting for evaluation for radiation treatment (RT) between 2012 and 2015 and who underwent mpMRI exams on a 3T Discovery MR750 (GE, Waukesha, WI, USA) were identified from the departmental database.

mpMRI exam consisted in part of: (i) axial T2w-MRI of the pelvis: resolution 1.25 mm × 1.25 mm × 2.5 mm; field of view: 320 mm × 320 mm; slice thickness = 2.5 mm (no gap); 72 slices, and (ii) DCE-MRI—12 series of T1 weighted (T1w) at 30–34 s temporal resolution, TR = 3.77–4.05 ms, TE = 1.69–1.78 ms acquired following intravenous bolus injection of a paramagnetic gadolinium chelate—0.1 mmol of gadobenate-glumine (Bracco Diagnostics Inc., Princeton, NJ, USA) per kilogram of body weight. The contrast is administered with a power injector (Spectris, Medrad Inc., Warrendale, PA, USA) at 2 mL/s and followed by a 20-mL saline flush.

The analysis pipeline is presented schematically in Figure 1. The DCE-MRI series were uploaded in MIM (MIM, Cleveland, OH, USA) and all subsequent analyses are performed using MATLAB R2014a (MathWorks, Natick, MA, USA) plugins. Prostate, PZ, and sample gluteus maximus (GM) volumes were manually contoured (Figure 1A). The individual steps of the analysis are described below:

Motion correction (Figure 1B). Each imaging set in the DCE-MRI series was aligned to the preceding one by finding the affine transformation (shifting, scaling, and rotation) that maximized their mutual information (22). Pixels within the prostate contour and extension by 18.75 mm (15 pixels) were considered. For each pair of before and after-correction images, the sum of squared pixel differences was computed and motion correction was carried out only when this sum was smaller post-correction.

For convenience, the feature map is represented using a 1 to 10 scale, grouping the W_wpⁱ F/ω in 10 bins. Feature range was bound by 5th and 95th percentile of the feature values in the entire population of subjects with positive biopsies. DCE-feature scores are allocated by uniformly dividing the bounded interval in ten regions. Pixels with feature values smaller than the lower bound or larger than the upper bound were assigned a DCE-feature-score of 1 and 10, respectively.

Spearman correlation coefficient (ρ) of the DCE-scores with the highest GS from the biopsy session was computed (patient level analysis). The GS were grouped in a clinically relevant categories, separating GS7 (3 + 4) and GS7 (4 + 3) (27). To study the ability of the features to discriminate between indolent (GS = 6) and aggressive (GS ≥ 7) lesions, the association between the quantitative features and these two classes was investigated. The significance of the median difference between groups was evaluated using Kruskal–Wallis test, followed by Dunn’s posttest. In addition, the AUC of the receiver operating characteristics (ROC) analysis was computed. A subset of the patients underwent MRI-ultrasound fused targeted (MRI-US) biopsies. The statistical test above were carried out for the DCE-scores of each biopsy location and biopsy GS (biopsy level analysis). Statistical analysis was performed using MATLAB. All tests were two-sided. Significance was set at a p-value <0.05.

The flow diagram of inclusion and exclusion criteria of the analyzed patients is presented in Figure S2 in Supplementary Material. Nine patients (9/71, 12%) were excluded because: four patients lacked complete DCE-MRI series and five had suboptimal contrast data due to acquisition artifacts or incomplete contrast administration. The clinical characteristics of the analyzed cohort are summarized in Table S1 in Supplementary Material. The highest GS from the biopsy session was used.

For each pixel in the dataset, DCE_N was aligned to DCE_N − 30, (N = 90, 120, …, 330) and motion was quantified as the mean absolute pixel gray scale-intensity difference between DCE_N and DCE_N − 30, before and after correction. For 27 patients (44%), the movement correction procedure resulted in essentially no change (change of pixels squared differences before and after: median = 0.02%; range = −2.38 to 1.51%). For 35 patients (56%), the procedure yielded significant improvement: median = 9.93%, range = 2.61–46.84%. The procedure is illustrated for the patient with the largest correction (Figure S3 in Supplementary Material).

All five segmentation methods resulted in significant correlations between the volumes of ROI_wp and GS. The Spearman correlation ρ of the volume of ROI_wp and GS (6, 3 + 4, 4 + 3, ≥8), its p-value, and the number of patients assigned to PZ or TZ, or rejected due to volumes less than 0.5 cc are presented in Table 2. While in general, with increasing β, the correlations with GS increased, the resulting smaller ROI_wp volumes caused a large number of patients to be rejected due to ROI_wp < 0.5 cc. β = 60 and 10% PZ vs TZ were selected. The median size of the well-perfused region ROI_wp across the patients was 1.95 cc (range 0.23–12.21 cc). Following 0.05 cc cleanup for disconnected small volumes, the resulting volumes ROI_wp were on average 18.19% smaller and ranged in size between 0.16 and 11.52 cc. Eight patients with ROI_wp <0.5 cc were eliminated from further processing. For the remaining 54 patients, the mean volume of ROI_wp was 2.57 cc (median = 1.84 cc, range 0.51–11.52 cc). Forty-one (76%) of the tumors were in the PZ (pzROI_wp). The mean volume of pzROI_wp was 1.08 cc (median = 0.80 cc, range 0.15–5.01 cc) and the mean volume of tzROI_wp was 3.09 cc (median = 1.28 cc, range 0.80–11.52 cc). The distributions of pzROI_wp volumes in association with GS divided in three (6, 3 + 4, >4 + 3) and four groups (6, 3 + 4, 4 + 3, 8 − 10) are shown in Figure 2A. In both cases, Spearman correlation coefficients were significant.

Six quantitative features (Table 1) were computed based on the signal S_ROI for the patients with suspicious area in PZ (pzROI_wp). The correlations between these features and the highest GS (6, 3 + 4 and >3 + 4) are presented before and after normalization with the signal from the muscle reference S_GM (Figures 2B,C). Note that the correlations with GS improve substantially following normalization with the muscle signal, reaching significance (p < 0.05) in all cases but one. When GS is divided in four groups: 6, 3 + 4, 4 + 3, 8 − 10 the same five features remained significant (Figure 2D). It should be noted that none of the features were significant for tzROI_wp (data not shown). In addition, comparisons including the distribution of the quantitative features in NAT_PZ and NAT_TZ we carried out (Figure S4 in Supplementary Material). Their corresponding AUC for discrimination between NAT_PZ and pzROI_wp ranged between 0.62 and 0.99. AUCs of DCE-scores in tzROI_wp were also high. In view of the small number of lesions in TZ, however, these results should be treated with caution.

Sixteen patients underwent MRI-US biopsies and had total of 35 positive biopsies (Figure S5 in Supplementary Material). After confirmation of the location of the actual biopsy, the region with needle tracks was back-projected onto T2w. The overlap between the biopsy region and the well-perfused region ROI_wp was denoted as usROI_wp. Quantitative features were computed for each biopsy in PZ (29/35) using the average signal in usROI_wp. The associations of the sixfeatures for PZ tumors and the targeted biopsy GS (divided in three and four groups) are presented in Figures 3A,B. The same four features were significant using three, four GS groups and when tumors were split into indolent (GS = 6) and aggressive (GS ≥ 7) [wash-in ratio reached marginal (p = 0.051) significance] (Figure 3C). The features resulted in AUC for classifying indolent vs aggressive PZ lesions ranging between 0.85 and 0.95 (Figure 3D).

Using W_wp and the DCE-score F from ROI_wp, each pixel in the prostate was scored for aggressiveness using 1 to 10 scale. In Figures 4A,B, two axial slices of the prostate are shown with indicated regions representing the intersection of the identified well-perfused region ROI_wp and the biopsy locations. A schematic representation of the prostate with the biopsy needle tracks (yellow) in the target volumes (1 in the green, 2 in the red, and 1 in the blue target) is depicted in Figure 4C. On pathology review, the biopsies were assigned GS = 6 (green target) and GS = 7 (for both biopsies in the red target). Map of aggressiveness was generated for early AUFC ratio feature. In Figures 4D,E, the map is overlaid on the patient axial MRIs slices. The areas of the positive biopsies are clearly identified. The color scheme of the maps is given on the right. There are no pixels with DCE-scores 9 and 10, indicating lack of areas with high GS. The highest DCE-score on the maps (left to right) was 6 and 8, consistent with the biopsy findings.

The MRI-US dataset was utilized to confirm the accuracy of the global map of aggressiveness. On one hand, the map of aggressiveness feature_Wwp was generated using only W_wp and F. On the other hand, the feature value for each biopsy ROI was estimated in the previous section using usROI_wp. In Figure S6 in Supplementary Material, the x-axis corresponds to the feature computed using usROI_wp. The y-axis is the mean value of the map of aggressiveness over the region of the biopsy, mean [feature_Wwp(mrusROI)]. There was an excellent concordance between these two estimates for each biopsy ROI (n = 29) for each feature: the correlation coefficients ranged between 0.85 and 0.99; the slope for the linear fit ranged between 0.87 and 0.98. This graph also confirms the co-localization of the “hot” spots of the map of aggressiveness with the biopsy location.