1.

Introduction

Prostate cancer is the second leading cause of cancer death in men in the U.S.^1,2 and the third leading cause of cancer death worldwide.³ Treatment of prostate cancer is dependent on cancer severity. High-grade cancer is treated with radical prostatectomy, radiation treatment, or hormone therapy, whereas low-grade cancer may be monitored by active surveillance.⁴ Ideally, noninvasive clinical imaging should allow physicians to estimate cancer grade in advance of confirmatory biopsy of suspicious regions.

Pathologic grading of prostate cancer has changed in the past decade by a consensus of the International Society of Urological Pathology (ISUP). Cribriform glands were upgraded from Gleason pattern 3 to 4 (G3 and G4) due to their association with more aggressive cancer.⁵ Moreover, since 2001, 16 papers plus many abstracts have shown the cribriform pattern to be associated with more biochemical failure and adverse pathology, such as stage, margin status, and cancer volume, compared to noncribriform (small fused) acini.⁶

Cribriform pattern is most frequently found in the peripheral zone (PZ). Furthermore, it is commonly known that $\sim 75\%$ of all types of prostate cancer occur in the PZ.^6,7 The PZ appears differently from the transition zone on MRI and therefore, different acquisitions are optimal for different zones in the PI-RADS v2 scoring system. Diffusion-weighted imaging (DWI) is weighted more heavily in the PZ while T2 is more relevant to the transition zone.

Recent research has shown that different magnetic resonance imaging (MRI) sequences have been able to discriminate cancer grades. Apparent diffusion coefficient (ADC) images derived from DWI-MRI have been shown to discriminate between the Gleason grades.^8–10 Additional MR sequences, such as T1, T2, MR spectroscopy, and dynamic contrast-enhanced imaging, add additional information for grade discrimination.^11–19

This study investigates various $b$-value combinations used to calculate ADC to determine combinations that optimally discriminate cancer grades, including cribriform versus noncribriform type, in both the entire prostate and limited to the PZ. We use a radiologic–pathologic (Rad-Path) correlation to validate various ADC measures in 1369 pathologically classified regions from 210 whole-mount processed prostate slides from 39 patients.

2.

Methods

2.3.

Tissue Preparation

After radical prostatectomy, prostate samples were fixed in formalin overnight and then sectioned using patient-specific 3-D printed slicing jigs designed to match axial MRI orientation (Fig. 1).²¹ Each 4-mm tissue slice was paraffin-embedded in large format cassettes. A microtome was then used to slice $10\mu \mathrm{m}$-thick sections, which were then transferred to whole-mount slides and hematoxylin and eosin stained. The slides were then digitized at $40\times$ magnification using an automated microscope²² (Nikon, Tokyo, Japan). Digitized slides were annotated by a board-certified urologic pathologist using current modified Gleason grading.^5,23,24 Regions of interest (ROIs) were manually drawn on the virtual slides creating annotated regions using different colors for: seminal vesicles, atrophy, high-grade prostatic intraepithelial neoplasia (HGPIN), G3, G4-fused small glands (G4FG), G4 cribriform glands (G4CG), and G5 carcinoma (Fig. 2). Regions not annotated were considered normal prostate. For subsequent analysis, normal prostate, seminal vesicles, atrophy, and HGPIN ROIs were combined and considered noncancerous (NC). Each digitized slide was then coregistered to the MRI and downsampled to corresponding voxel resolution (Fig. 3) using methods and software reported previously.^21,22,25,26 Control point nonlinear coregistration was used to nonlinearly warp the histology and contoured annotation to match the endorectal coil compressed prostate in the T2-weighted MRI using the imwarp command in MATLAB (Mathworks Inc.).

Fig. 1

Demonstration of the use of patient-specific 3-D printed prostate slicing jig for sectioning tissue to match the orientation of anatomical presurgical MRI. The posterior surface of the prostate most distorted by the endorectal coil is oriented upward toward the mold insertion opening.

Fig. 2

Examples of representative regions annotated using the Gleason system by our pathologist at $40\times$ magnification.

Fig. 3

Demonstration of the (a) whole mount histology annotated by our urologic pathologist (b) coregistered to the T2 anatomic imaging using a control-point coregistration (c) with edge detection and applied warping to annotation and histology.

3.

Results

A total of 1369 annotated pathological regions from 210 prostate slices from 39 patients were included in the analysis of the entire prostate. This included 210 regions considered normal prostate tissue, 25 of seminal vesicles, 692 atrophies, 77 HGPIN, 252 G3, 81 G4FG, 24 G4CG, and 8 G5. When limited to the PZ, a total of 802 annotated pathological regions including 207 regions considered normal prostate tissue, 9 of seminal vesicles, 376 atrophies, 23 HGPIN, 127 G3, 38 G4FG, 16 G4CG, and 6 G5. Regions not annotated by our pathologist were considered normal.

ADC alone significantly discriminates high-grade (G4+) cancer from low-grade (G3) and other normal tissue. The optimal $b$-value combinations and thresholds are shown in Table 2 (PZ) and Table 3 (whole prostate) marked by an asterisk. Figure 4 shows the ROC curves for G4+ versus G3 and NC, G4+ versus NC, G4CG versus NC, G4CG versus G3 and NC, and G4CG versus G3 within the PZ. Figure 6 shows the ROC curves for each combination performed in the whole prostate. Each $b$-value combination is plotted separately with the optimal combination highlighted by a thicker line.

Table 2

Analysis of the PZ. Summary table of ADC b-value combinations paired with the respective AUC, cut-off, and means and medians for each group being compared within the PZ of the prostate.

Table 3

PZ and TZ/CZ combined. Analysis of the entire prostate. Summary table of ADC b-value combinations paired with the respective AUC, cut-off, and means and medians for each group being compared over the entire prostate. HG = high grade, LG = low grade.

Fig. 4

ROC curves plotted for each $b$-value combination against five different comparisons (HG versus G3 + noncancerous tissue, HG versus LG, cribriform glands versus NC tissue, cribriform glands versus LG, and cribriform glands versus G3 + NC tissue). The bold line represents the $b$-value combination with the highest AUC. AUC values are shown in Table 2. TPR, true-positive rate and FPR, false-positive rate. (a) G4CG versus G3, (b) G4CG versus NC, (c) G4CG versus G3 and NC, (d) G4+ versus G3, and (e) G4+ versus G3 and NC.

Results from the ROC analysis indicate ADC significantly discriminated high-grade cancer from low-grade and NC tissue. Figure 4 marks significant comparisons with an asterisk and Table 2 reports $p$ values. Figure 5 shows that in general, diffusion is more restricted in higher grade cancer. However, comparison of G5 and G4CG or of G4CG and G4FG were not significant. Mean difference was significant between G5 and G3 using the $b$-value 0 to 200; G4CG and G3 using 0 to 2000, 100 to 2000, 1000 to 2000, 200 to 2000, 50 to 2000, and 500 to 2000; G4FG and G3 using 50 to 2000, 50 to 1000, 200 to 2000, 200 to 1000, 100 to 2000, 100 to 1000, 0 to 500, 0 to 2000, and 0 to 1000; and overall combinations between all high-grade cancers and NC tissue. The whole prostate analysis, including ROC curves from additional comparisons, is shown in Fig. 6 and Table 3, G5 versus G3 cancer versus noncancer (G3+ versus NC), G4FG versus NCG5 versus NC, G4FG versus G3 and NC, G5 versus G3 and NC, and G4FG versus G3.

Fig. 5

Bar charts indicating the mean ADC for each $b$-value combination within 1369 pathologist annotated regions including NC tissue, Gleason grade 3 (G3), grade 4 fused glands (G4FG), grade 4 cribriform (G4CG), and grade 5 (G5).

Fig. 6

ROC curves plotted for each $b$-value combination against six different comparisons (HG versus G3 + NC tissue, HG versus LG, cancerous tissue (G3+) versus NC tissue, fused glands versus NC tissue, cribriform glands versus NC tissue, and grade 5 versus NC tissue). The bold line represents the $b$-value combination with the highest AUC. AUC values are shown in Table 3. (TPR, true-positive rate and FPR, false-positive rate.)

The strongest $b$-value combination that was clinically relevant to discriminate high- from low-grade and NC tissue was 50 to 2000. For the splitting of individual Gleason grades, there was not a clear overall best combination, but inclusion of the highest $b$-value (e.g., 1000 to 2000, 50 to 2000, 0 to 2000) had larger AUCs. Specific values for each grade are shown in Table 2 for the PZ and Table 3 for the whole prostate, along with optimal cut-offs for all $b$-value combinations.

4.

Discussion

This study used Rad-Path correlation to assess the sensitivity and specificity of various $b$-value combinations for calculating ADC to discriminate aggressive high-grade prostate cancer from indolent cancer. We found that in the PZ, the optimal $b$-value combination to discriminate high- from low-grade cancer and NC tissue was 50 to 2000, which resulted in an AUC of 0.813 with an optimal ADC threshold of $1.042\times {10}^{-3}{\mathrm{mm}}^2/\mathrm{s}$. Commonly acquired $b$-value combinations, such as 0 to 1000 and 0 to 2000, performed only marginally worse with AUC’s of 0.799 and 0.812, respectively.

Previous studies have shown that ADC is inversely correlated with cancer grade.^8,11,27,28 These previous reports have grouped G4CG, G4FG, and G5 into one class of high-grade. This study divided these two grades into three subgroups to investigate the different diffusion properties of these three subgroups. Considering that cribriform pattern is more aggressive than G4FG, it was important to investigate our ability to identify it on MRI.

We found that mean ADC values for G4FG, G4CG, and G5 are not significantly different from one another although all three were significantly different from low-grade cancer in the higher performing $b$-value combinations (Fig. 5). Other studies have shown the same difficulty in separating G4 from G5, and therefore, merge G4 and G5 in their analysis.^8,11 We found ADC alone can be used to discriminate each of these high-grade cancer types from all other prostate tissue types using $b$-value combinations. Importantly, cribriform glands hold a greater predictive value for biochemical failure than small fused glands.²⁹ Our data show that lower ADC scores correlate with more severe cancer, with the exceptions reported above including G4CG not showing this perfect inverse correlation. The combinations most responsible for a variation from this inverse correlation are 0 to 500 and 1000 to 2000, where both group means and medians are altered, the former raising G4 cribriform, and the latter raising G5. This may be attributed to the diffusion compartments highlighted by these $b$-value combinations. More research is necessary to determine why the trend reverses.

We found that the inclusion of higher b-values in the ADC calculation resulted in increased sensitivity and specificity for differentiating high-grade cancer, which is supported by other work.^20,28 The pathology of G5 is very different from the pathology of G4CG.^23,24,30 G5 is characterized by infiltrating single cells, whereas G4CG has distended gland spaces with multiple lumens. A possible explanation for the variance in diffusion between these morphologies is that water may be able to diffuse through the stroma surrounding single cells more easily than through the highly cellular, thick-walled epithelium of cribriform glands.

There are several sources of potential error in our analysis. We chose to utilize an open design for the 3-D printed slicing jigs for ease of use. We initially experimented with a design similar to Trivedi et al.³¹ but found our pathology assistants preferred the design shown in Fig. 1 due to ease of use and minimal workflow disruption. Prostate samples fit within the mold snugly as the 3-D renderings used for the design were deflated subtly. It is plausible that small slice orientation discrepancies occurred. Future studies should intentionally perturb the coregistration scheme to determine downstream cancer classification effects. It will also be beneficial to the field to explore a more sophisticated fitting of the diffusion signal, and how the resulting maps effect cancer discrimination. An additional analysis of the intra- and intersubject variability of tumor ADC values would be an interesting direction for future research.

This study disclosed that while $b$-value selection to discriminate high-grade prostate cancer can be optimized, most common combinations perform suitably. ADC alone can be used to discriminate individual high-grade cancer types from low-grade cancer and NC tissue to a significant degree, with specific combinations outperforming others. We expect this information will guide clinical interpretation of ADC maps and spur additional research.