Histological validation of prostate tissue composition measurement using hybrid multi-dimensional MRI: agreement with pathologists’ measures

Abstract

Purpose

To validate prostate tissue composition measured using hybrid multi-dimensional MRI (HM-MRI) by comparing with reference standard (ground truth) results from pathologists’ interpretation of clinical histopathology slides following whole mount prostatectomy.

Materials and methods

36 prospective participants with biopsy-confirmed prostate cancer underwent 3 T MRI prior to radical prostatectomy. Axial HM-MRI was acquired with all combinations of echo times of 57, 70, 150, 200 ms and b-values of 0, 150, 750, 1500 s/mm² and data were fitted using a 3-compartment signal model using custom software to generate volumes for each tissue component (stroma, epithelium, lumen). Three experienced genitourinary pathologists independently as well as in consensus reviewed each histology image and provide an estimate of percentage of epithelium and lumen for regions-of-interest corresponding to MRI (n = 165; 64 prostate cancers and 101 benign tissue). Agreement statistics using total deviation index (TDI_0.9) was performed for tissue composition measured using HM-MRI and reference standard results from pathologists’ consensus.

Results

Based on the initial results showing typical variation among pathologists TDI_0.9 = 25%, we determined we will declare acceptable agreement if the 95% one-sided upper confident limit of TDI_0.9 is less than 30%. The results of tissue composition measurement from HM-MRI compared to ground truth results from the consensus of 3 pathologists, reveal that ninety percent of absolute paired differences (TDI_0.9) were within 18.8% and 22.4% in measuring epithelium and lumen, respectively. We are 95% confident that 90% of absolute paired differences were within 20.6% and 24.2% in measuring epithelium and lumen, respectively. These were less than our criterion of 30% and inter-pathologists’ agreement (22.3% for epithelium and 24.2% for lumen) and therefore we accept the agreement performance of HM-MRI. The results revealed excellent area under the ROC curve for differentiating cancer from benign tissue based on epithelium (HM-MRI: 0.87, pathologists: 0.97) and lumen volume (HM-MRI: 0.85, pathologists: 0.77).

Conclusion

The agreement in tissue composition measurement using hybrid multidimensional MRI and consensus of pathologists is on par with the inter-raters (pathologists) agreement.

Graphical Abstract

Introduction

Prostatic problems such as prostate cancer and benign prostatic hyperplasia are very common among aging men, causing substantial suffering and burden to the US healthcare system [1]. While histological changes remain the reference standard for diagnosis of prostatic problems, multiparametric MRI (mpMRI) is increasingly being used for imaging the prostate. The assessment of the prostate using mpMRI is through T2-weighted, diffusion-weighted, and dynamic contrast enhancement MR images [2]. However, up to 30% of clinically significant prostate cancers can be missed on mpMRI [3], and therefore histological changes using invasive biopsies samples remains the gold standard for prostate cancer diagnosis. While conventional mpMRI does not reveal these histological changes, MRI microscopy of fixed prostate tissue using a preclinical 16 T MR has been used to image individual prostatic glands at sufficient resolution (40 μm isotropic) and demonstrated the different MR properties of gland microstructures and changes in these microstructures associated with these prostatic conditions [4, 5]. However, the spatial resolution of clinical MRI scans is far too low to resolve glandular structures as prostate acini are approximately 0.1 mm in diameter [6]. Therefore, newer MRI techniques [7] such luminal water imaging [8], vascular extracellular and restricted diffusion for cytometry in tumors (VERDICT) [9], restriction spectrum imaging [10], and hybrid multidimensional MRI (HM-MRI) [11] have been developed to measure these tissue microstructures non-invasively.

Hybrid Multidimensional MRI (HM-MRI) detects tissue heterogeneity at the microscopic level through the combination of the most effective MRI contrast mechanisms that are currently used clinically—T2 relaxometery and diffusion (T2 and Apparent Diffusion Coefficient or ADC measurements) to obtain quantitative tissue microstructure measurements, specifically tissue composition maps: fractional volumes of lumen, stroma, and epithelium. Recent studies have shown that these volume fractions change with difference prostatic pathology type, such as with the presence of prostate cancer [12] and the associated Gleason grade [13]. HM-MRI measures the change in ADC and T2 as a function of echo time (TE) and diffusion weighting (b-value), respectively [14–16], and exploits the distinct physical properties: the coupled T2 and ADC values associated with each of these tissue components [4] to measure tissue composition changes non-invasively and diagnose prostate cancer. Our recent feasibility study [11] showed that prostate tissue composition can be measured non-invasively using HM-MRI and has the potential to improve prostate cancer diagnosis and determine its aggressiveness. However, our previous study on measurements of prostate tissue composition using HM-MRI [11] lacked validation with ground truth histopathology results. A more recent validation study using quantitative histology of 75 prostatic sections from prostatectomy samples was able to demonstrate that the correlation between tissue measures from HM-MRI and histology is excellent (r ~ 0.9) [17]. However, quantitative histology is not considered the ground truth for clinical purposes, and therefore studies comparing HM-MRI results with expert pathologists are needed. Therefore, the purpose of this study is to validate prostate tissue composition measured using HM-MRI by comparing with reference standard results from pathologists’ interpretation of clinical histopathology slides following whole mount prostatectomy in a larger statistically powered study.

Materials and methods

Study participants

This study involved retrospective analysis of prospectively collected data. This study was approved by the institutional review board and was in compliance with the Health Insurance Portability and Accountability Act. Participants had provided informed written consent.

Thirty-six consenting participants with elevated prostate specific antigen and prior histologically confirmed prostate cancer through prior biopsy that underwent prostate MRI including Hybrid Multi-dimensional MRI followed by subsequent radical prostatectomy were recruited for this study at our research center between December 2016 and April 2019. Individuals who were treated with radiation or hormonal replacement therapy (leading to alterations in prostatic signal on MRI) before prostatectomy or MRI were not eligible for inclusion.

MRI acquisition

Participants underwent preoperative multiparametric MRI with a 3 T Philips Ingenia or Achieva MR scanner using a combination of a six-channel phased array coil placed around the pelvis and an endorectal coil (Medrad, Bayer Healthcare). The HM-MRI sequence consists of a spin-echo module with diffusion sensitizing gradients placed symmetrically about the 180° pulse followed by single shot echoplanar imaging readout. Axial images using HM-MRI were acquired with all combinations of TE = 57, 70, 150, 200 ms and b-values of 0, 150, 750, 1500 s/mm², which resulted in a 4 × 4 array of data associated with each voxel. Axial HM-MRI images were oriented perpendicular to the rectal wall, as guided by sagittal images to match the alignment of MR images more closely with whole mount histology sections. Fat saturation was performed using spectrally selective adiabatic inversion recovery. The MRI parameters were the following: in-plane resolution, 1.5 × 1.5 mm²; scan matrix, 120 × 120; field of view, 180 × 180 mm²; repetition time, 5 s; number of slices, 18; slice thickness, 3 mm; and reconstruction matrix, 128 × 128. The acquisition time was 10–12 min. In addition, standard clinical multi-parametric MRI scans including T2-weighted (axial, coronal, sagittal), diffusion-weighted imaging (axial), dynamic contrast enhanced imaging (axial) were also performed using the protocol described in our previously published work [18–20].

Histopathology

Participants subsequently underwent radical prostatectomy and the excised prostate was fixed in formalin and serially sectioned in approximately the same plane as the MR images. Whole mount tissue sections were stained with hematoxylin and eosin (H&E) to create histological slides. The slides were evaluated for prostatic adenocarcinoma by expert genitourinary pathologists as per standard clinical procedure. Areas of tumor were marked on the histologic slides. The slides were scanned at ×20 magnification using Olympus VS120 whole mount digital microscope (Olympus Corporation, Waltham, MA) and saved as Olympus Virtual Slide images. Histology and MRI images were co-registered visually by the consensus of an expert radiologist (AO, 15 + years of experience with prostate MRI), pathologist (TA, 15 years of experience with genitourinary pathology), and medical physicist (AC, 8 years of experience with prostate MRI and pathology). Regions-of-interest (ROI) including cancers and benign tissue from peripheral, central, and transition zones were marked on histology and MR images (ADC maps) on sites of prostatectomy verified malignancy and benign tissue from peripheral, central, and transition zones. The regions-of-interest were defined on pathology images while the researchers were blinded to MRI results. The minimum size criteria was 5 mm × 5 mm. These regions-of-interest (ROIs) on pathology slides were converted from Virtual Slide Image (.vsi) file format (Olympus proprietary format) to commonly used Tagged Image File Format (TIFF) using BIOP VSI reader plugin on ImageJ (National Institutes of Health, Bethesda, MD). The TIFF image file format is based on the lossless compression method, so that an image is stored with loss of less than 1% of the original data. These images can be opened on a computer using Windows Photo Viewer (image viewer found on Windows computers). The data were saved in a secure share drive—Box Business Plus that the pathologist used to access images. This provides a secure encrypted storage platform with ability to view images and track user activity.

Pathology based tissue composition

Three experienced board certified genitourinary pathologists (15, 6, 11 years of experience with prostate pathology) independently reviewed each image ROI visually and provided an estimate of percentage of epithelium and lumen for each ROI. Since the sum of the tissue components: stroma, epithelium, and lumen is 100%, pathologists were not tasked with estimating stromal volume. An initial review of 10 patients was done to obtain preliminary results needed for sample size and acceptance criterion determination. This was followed by independent analysis of the whole data set by each of the pathologists. Four weeks after the initial review, pathologists repeated the review of the initial 10 cases and recorded their results. This allowed us to calculate intra-observer variation for these tissue percentage estimates. Upon completion of the independent review by each pathologist, the three pathologists reviewed the ROIs in consensus for the estimates of epithelium and lumen and recorded the results. These results served as the ground truth measures for this cohort.

MRI based tissue composition

The analysis was performed on a voxel-by-voxel basis using a custom software written in MATLAB (MathWorks, Natick, MA). The prostate tissue component volumes were calculated by compartmental analysis of HM-MRI data. The HM-MRI signals were modeled as unmixed pools of water in three tissue components: stroma, epithelium, and lumen each with distinct and coupled T2 and ADC properties as explained in our previously published articles [11, 17]. The fractional volumes of tissue component were calculated on a voxel-by-voxel basis by fitting the following equation using Nonlinear Least Squares method with an in-house MATLAB (MathWorks, Natick, MA) program:

$$\frac{S}{S_{\text{o}}}=\sum _{n=1}^{n=3}{V}_n\times \exp \left(-AD{C}_n\times b-\frac{TE}{T{2}_n}\right)$$

where V_n, T2_n, and ADC_n are the volume fractions, T2_n and ADC_n values for each tissue compartments (stroma, epithelium, and lumen) and S is the signal intensity at each combination of echo time and b-value and S₀ is the signal intensity at the lowest echo time and b-value. The sum of volumes of the three tissue components = 100%.

The same regions-of-interest taken on pathology images on pathology confirmed cancerous and benign tissue were then taken on HM-MRI results: tissue composition maps to calculate mean values of epithelium and lumen volume. These results were then correlated with pathologists’ results and measure agreement with reference values from pathologists using whole mount prostatectomy specimens.

Statistical analysis

We compared tissue components volumes estimated from HM-MRI with independent measure by pathologists and the consensus of 3 expert pathologists. Intra and inter-observer variation between each pathologist was calculated. The following agreement statistics were performed by an expert biostatistician (LL, 40 years of experience) as detailed in Lin et al. [21].

Concordance Correlation Coefficient (CCC) – CCC is for measuring agreement between observations and target values. It is defined as the product of accuracy and precision coefficient. CCC = 1 represents a perfect agreement (all observations = target values). CCC = 0 represents no agreement. CCC = −1 represents a perfect reverse agreement.

Accuracy—It is the measures the closeness of the observations and target values in terms of both means and variances (marginal distributions). No differences in means and variances occur when accuracy coefficient = 1.

Precision—It is the Pearson correlation coefficient measuring how closely the observations deviate from the best-fit linear line.

Total Deviation index—Total deviation index (TDI(1-p)) translates the mean of the squared difference between observations and target values into an index that can be directly compared to a predetermined criterion. The TDI(1-p) describes a boundary such that a majority, 100(1-p) per cent, of the observations are within the boundary (measurement unit and/or per cent) from their target values.

TDI_0.9 – 90% quantile of the absolute differences of the paired readings. In other words, when TDI_0.9 = 15, it means that 90% of paired observations are within 15% of each other. Note that our measuring unit is %. This is the same as the total analytical error as suggested by CLSI-EP21 (2016) and CLSI-EP21-A (2003) guidance [22, 23]. The 95% upper confidence limit of TDI is the same as using the tolerance interval as suggested by the CLSI-EP21 and CLSI-EP21-A guidance. In CLSI-EP21-A, the two-sided interval of the paired differences is used. Here, we use the absolute difference instead of the difference, and therefore we use the 95% one-sided upper confidence limit of the absolute paired differences. This 95% upper confidence limit of the TDI_0.9 means that we are 95% confident that 90% of the absolute paired differences are within TDI_0.9.

Note that CCC, accuracy, and precision coefficients are highly dependent on the study range, while TDI does not. For this study, the data ranges of measuring lumen (%) were shorter than that of measuring epithelium (%) and therefore would yield lower CCC, accuracy, and precision coefficients. Therefore, TDI measures, which also conforms to the above CLSI guidance, was used as the primary statistical endpoint.

Receiver operating characteristic (ROC) analysis was used to evaluate the performance of fractional volume of tissue components in differentiating cancer from benign prostatic tissue measured using HM-MRI and pathologists’ measures.

Our hypothesis and study endpoint is that agreement in tissue composition measurement using our new non-invasive technique—Hybrid multi-dimensional MRI and consensus of pathologists is on par with independent pathologist’s evaluation or inter-raters (pathologists) agreement. Furthermore, these tissue measures from HM-MRI can potentially be used for non-invasive prostate cancer diagnosis.

Results

Participant and tumor characteristics

The mean age of all included men was 60 years (range of 47–70 years), and mean PSA was 9.0 ng/mL (range of 2.5–64.0 ng/mL). The median time between MRI and prostatectomy was 20 days (range of 1–34 days). A total of 64 tumor ROIs (11 Gleason 3 + 3, 38 Gleason 3 + 4, 11 Gleason 4 + 3, 1 Gleason 4 + 4, 3 Gleason 4 + 5) and 101 benign tissue ROIs (34 from peripheral zone, 32 from central zone, and 35 from the transition zone) were included in the analysis. The ROI size ranged from 44 to 1269 mm² with a mean of 199.5 ± 142.6 mm2. Due to extensive cancer in the prostatectomy specimens in some cases, benign tissue from certain zones were not included and hence the lower number of benign ROIs included in the study than the number of patients.

Initial results for sample size determination

At first, we performed an initial study to obtain 38 samples including both prostate cancer and benign tissue from 10 patients. For each sample, fraction volumes of the epithelium and lumen were measured by the custom software using HM-MRI data and by three pathologists. Table 1 presents the agreement statistics among the three pathologists and between HM-MRI and the mean of three pathologists. These results show that the agreement between HM-MRI and the mean of pathologists was in between the agreement among individual pathologists. Based on CLSI-EP21 and CLSI-EP21-A guidance, we used TDI_0.9 for the sample size calculation. Because we will interpret the results based on the final data collection, we simply discuss the TDI results here.

Table 1.

Agreement statistics among three pathologists and between HM-MRI and mean of three pathologists

Area	Comparisona	CCC	Precision coefficient	Accuracy coefficient	TDI0.9b
Epithelium	P1 vs P2	0.9328	0.9728	0.9589	14.45 (17.07)
	P1 vs P3	0.7618	0.9402	0.8103	24.37 (29.22)
	P2 vs P3	0.8760	0.9515	0.9206	15.60 (18.93)
	HM-MRI vs P	0.7707	0.8404	0.9171	20.82 (25.15)
Lumen	P1 vs P2	0.6835	0.7806	0.8756	18.62 (22.48)
	P1 vs P3	0.3926	0.6665	0.5891	27.71 (32.80)
	P2 vs P3	0.5478	0.6797	0.8060	17.34 (20.78)
	HM-MRI vs P	0.6385	0.7959	0.8022	17.99 (21.25)

^aP1, P2, P3 are pathologists 1, 2, and 3; and P is the mean of 3 pathologists

^bShown in parenthesis is the 95% one-sided upper confidence limit

1. For measuring fractional volume of epithelium: 90% of the absolute paired differences between HM-MRI and the mean of three pathologists were within 20.8%. On the other hand, pairwise absolute differences among the three individual pathologists were within 14.5%, 24.4%, and 15.6%.

2. For measuring fractional volume of lumen: 90% of the absolute paired differences between HM-MRI and the mean of three pathologists were within 18.0%. On the other hand, pairwise absolute differences among the three individual pathologists were within 18.6%, 27.7%, and 17.3%.

Sample size and acceptance criterion

Based on the above initial results, the biostatistician determined using TDI_0.9 = 25% as our typical variation among pathologists in measuring epithelium and lumen, and we allowed for 5% cushion (acceptable TDI_0.9 = 30%, as this is close to the upper limit of TDI between pathologists), the sample size needed is n = 165 for the one-sided alpha = 0.05 and power = 0.95, using Eq. 4.3 (Lin et al. 2012, p. 72) [21]. Here, we assume samples as independent, even though each patient can have 3–4 samples, as these come from distinct lesion or anatomic zone with distinct pathology and can be considered independent samples. We will declare acceptable agreement if the 95% one-sided upper confident limit of TDI_0.9 is less than 30%. In other words, if we are 95% confident that 90% of absolute pair differences between HM-MRI and the consensus scores of pathologists are within 30%, we will accept the agreement.

Final results

We obtained 165 samples from 36 patients. These included the initial 38 samples from the pilot study. For each sample, fractional volumes of epithelium and lumen (%) were measured by HM-MRI, three pathologists, and the consensus of the three pathologists. Also, from the initial 38 samples collected in the pilot study, we obtained duplicate measures from each pathologist to evaluate the intra-rater agreement.

The data appeared to be well behaved and we did not observe any obvious outliers. Also, the distribution of the paired differences did not appear to deviate much from the normality assumption. The above will be clearly shown in Figs. 1 and 2 below. According to Lin et al. [24], the parametric TDI method performs almost flawlessly even when n = 20 under Poisson and uniform distributions, and it has higher power than the non-parametric method. The non-parametric method should be used when there are obvious outliers and/or obvious deviations from the normality assumption. For this study, the parametric TDI was used.

Fig. 1.

Intra-rater agreement plots in measuring epithelium (%) and lumen (%)

Fig. 2.

Inter-Rater and between methods agreement plots in measuring epithelium (EPI%) and lumen (lumen%)

Within each pathologist (intra-rater)

Table 2 presents the intra-rater agreement statistics for each pathologist in measuring epithelium (%) and lumen (%). Figure 1 presents the respective intra-rater agreement plots. Compared to the Bland and Altman plot [25], this type of agreement plot shows the real data values with more direct and intuitive information about accuracy and precision.

Table 2.

Intra-rater agreement statistics for each pathologist

Area	Pathologist	CCC	Precision coefficient	Accuracy coefficient	TDI0.9a
Epithelium	1	0.9526	0.9543	0.9982	12.61 (15.31)
	2	0.9846	0.9911	0.9934	6.64 (7.99)
	3	0.8858	0.9577	0.9249	13.11 (15.36)
Lumen	1	0.8700	0.8803	0.9883	13.79 (16.73)
	2	0.8661	0.9383	0.9230	9.46 (11.16)
	3	0.5799	0.6236	0.9299	14.31 (17.33)

^aShown in parenthesis is the 95% one-sided upper confidence limit

Pathologist 2 yielded the best results in measuring epithelium (%) and lumen (%) with CCC = 0.984 and 0.866, respectively. The accuracy and precision coefficients were excellent for measuring epithelium (> 0.99) and lumen (> 0.92). The data range of measuring lumen was shorter than that of measuring epithelium and therefore had lower CCC, precision, and accuracy coefficients. Ninety percent of absolute paired differences were within 6.64% and 9.46% in measuring epithelium and lumen, respectively. We are 95% confident that 90% of absolute paired differences were within 7.99% and 11.16% in measuring epithelium and lumen, respectively.

Pathologists 1 and 3 yielded similar TDI_0.9 results in measuring epithelium (~ 13%) and lumen (~ 14%). Pathologist 3 had slightly lower first (earlier) epithelium measurements than that of second (later) measurements and resulted in a lower accuracy coefficient (0.925) and CCC (0.886). The lower CCC, precision, and accuracy coefficients in measuring lumen were due to a smaller data range. Pathologist 1 had an excellent accuracy coefficient in measuring epithelium (0.998) and lumen (0.988), but with less precision compared to Pathologist 2.

Among three pathologists (inter-raters) and between methods

A representative example is provided in Fig. 3. Table 3 presents the final detailed summary of prostate tissue composition measures from MRI, individual assessment of 3 pathologists and consensus of the 3 pathologists with mean fractional volume and standard deviations reported. Tissue composition measured using HM-MRI and pathologists’ consensus were similar for epithelium 28.3 ± 12.8 vs 27.7 ± 18.0% and lumen volume 26.3 ± 12.8 vs 36.6 ± 10.0%. HM-MRI measures lie within the range of individual measures from each of the pathologists.

Fig. 3.

Representative images from a 65-year-old man with prostate specific antigen level of 64 ng/ml with confirmed Gleason 4 + 5 cancer in the right peripheral zone in mid prostate. The representation image shows analysis performed (A) showing T2-weighted image (top left), apparent diffusion (ADC) coefficient map at TE = 57 ms (bottom left), tissue composition maps for epithelium (top right), and lumen (bottom right). The region of interest in the right peripheral zone is Gleason 4 + 5 tissue. The tissue composition for this tissue using MRI was epithelium = 72.6% and lumen = 8.2%. While analysis of histology images (C) taken from whole mount prostatectomy specimen (B) revealed a consensus measure of epithelium = 80% and lumen 5%. Each individual pathologist reported the following measures (P1: epithelium 85% lumen 5%, P2: epithelium 81% lumen 9%, P3: epithelium 65% lumen 5%)

Table 3.

Summary of prostate tissue composition measures from MRI, individual assessment of 3 pathologists and consensus of the 3 patholo- gists mean fractional volume ± standard deviation (%) reported

	HM-MRI	Consensus of 3 pathologists	Pathologist 1	Pathologist 2	Pathologist 3
Epithelium	28.3 ± 12.8	27.7 ± 18.0	30.2 ± 18.8	21.7 ± 17.4	20.6 ± 12.9
Lumen	26.3 ± 12.8	36.6 ± 10.0	38.8 ± 12.7	29.4 ± 10.0	28.9 ± 6.8

Table 4 presents the inter-rater agreement statistics for pairwise comparisons among three pathologists, and for comparing HM-MRI the consensus of three pathologists in measuring epithelium and lumen. Figure 2 presents the respective agreement plots in measuring epithelium (EPI%) and lumen (Lumen%).

Table 4.

Pairwise agreement statistics among three pathologists and between methods

Area	Comparisona	CCC	Precision coefficient	Accuracy coefficient	THI0.9b
Epithe-lium	P1 vs P2	0.8445	0.9411	0.8974	17.53 (18.80)
	P1 vs P3	0.7403	0.9355	0.7913	20.73 (22.33)
	P2 vs P3	0.8948	0.9385	0.9534	11.57 (12.67)
	HM-MRI vs P	0.7325	0.7767	0.9431	18.79 (20.58)
Lumen	P1 vs P2	0.5547	0.7637	0.7263	20.55 (22.15)
	P1 vs P3	0.3902	0.6870	0.5680	22.42 (24.23)
	P2 vs P3	0.6325	0.6809	0.9288	12.12 (13.28)
	HM-MRI vs P	0.4985	0.7203	0.6921	22.42 (24.17)

^aP1, P2, P3 are pathologists 1, 2, and 3; and P is the consensus of 3 pathologists

^bShown in parenthesis is the 95% one-sided upper confidence limit

Pathologist 1 yielded higher results in measuring epithelium (%) and lumen (%) than those of the other two pathologists and therefore yielding lower accuracy coefficients with good precision coefficients. Again, the data range of measuring lumen was shorter than that of measuring epithelium and therefore had lower CCC, precision, and accuracy coefficients. For comparing pathologists 1 vs 2, 90% of absolute paired differences were within 17.5% and 20.6% in measuring epithelium and lumen, respectively. We are 95% confident that 90% of absolute paired differences were within 18.8% and 22.2% in measuring epithelium and lumen, respectively. For comparing pathologists 1 vs 3, 90% of absolute paired differences were within 20.7% and 22.4% in measuring epithelium and lumen, respectively. We are 95% confident that 90% of absolute paired differences were within 22.3% and 24.2% in measuring epithelium and lumen, respectively.

The agreement between pathologists 2 and 3 was much higher with CCC = 0.894 and 0.632, respectively, with good accuracy and precision. Ninety percent of absolute paired differences were within 11.6% and 12.1% in measuring epithelium and lumen, respectively. We are 95% confident that 90% of absolute paired differences were within 12.7% and 13.3% in measuring epithelium and lumen, respectively.

Agreement between HM-MRI and the consensus of three pathologists had similar accuracy as the pathologists 2 vs 3, but with lower precision in measuring epithelium. The between methods agreement statistics were in between those seen with comparing pathologists 1 vs 2 and comparing pathologists 1 vs 3.

The results of tissue composition measurement from MRI compared to reference standard results from the consensus of 3 pathologists, reveal that ninety percent of absolute paired differences were within 18.8% and 22.4% in measuring epithelium and lumen, respectively. We are 95% confident that 90% of absolute paired differences were within 20.6% and 24.2% in measuring epithelium and lumen, respectively. These were less than our criterion of 30% and therefore we accept the agreement performance of HM-MRI. We also conclude that the between methods agreement is on par with the inter-raters (pathologists) agreement.

Cancer diagnosis

The epithelium volume in cancer was 38.4 ± 12.8% using HM-MRI and 45.4 ± 16.5% based on pathologists’ consensus. The lumen volume in cancer was 17.8 ± 10.4% using HM-MRI and 31.1 ± 9.3% based on pathologists’ consensus. The epithelium volume in benign tissue was 22.0 ± 7.7% using HM-MRI and 16.5 ± 5.7% based on pathologists’ consensus. The lumen volume in cancer was 31.7 ± 11.2% using HM-MRI and 40.0 ± 8.8% based on pathologists’ consensus.

Area under the ROC curve for differentiating prostate cancer from benign tissue based on epithelium volume measured using HM-MRI was 0.870 (95% confidence interval of 0.808–0.932, p < 0.05) and pathologists 0.969 (95% confidence interval of 0.946–0.991, p < 0.05). Using lumen volume yielded similarly excellent AUC using either HM-MRI 0.847 (95% confidence interval of 0.781–0.913, p < 0.05) or pathologists measures 0.768 (95% confidence interval of 0.692–0.845, p < 0.05).

Discussion

The results of this study demonstrate that the tissue composition measured non-invasively using hybrid multidimensional MRI matches very closely with the reference standard results from the consensus of 3 expert pathologists. The results reveal that ninety percent of absolute paired differences (TDI_0.9) were within 18.8% and 22.4% in measuring epithelium and lumen, respectively. We are 95% confident that 90% of absolute paired differences were within 20.6% and 24.2% in measuring epithelium and lumen, respectively. These were less than our criterion of 30% and therefore we accept the agreement performance of HM-MRI. In addition, the between methods agreement statistics (HM-MRI vs consensus) were in between those seen with comparing agreement between pathologists. Therefore, we conclude that the between methods agreement is on par with the inter-raters (pathologists) agreement. In addition, high diagnostic accuracy based on high area under the receiver operating characteristic curve using tissue composition measures from HM-MRI and pathologists suggests prostate cancer could be differentiated from benign tissue as indicated by increased epithelium and reduced luminal volume in prostate cancer compared to benign tissue.

Several previous studies have estimated tissue composition. The tissue composition estimated non-invasively in this study using both HM-MRI and pathologists are similar to the tissue composition: fractional volumes of epithelium and lumen reported in the literature from morphometric analysis of H&E stained prostate tissue [8, 12, 13, 26]. The tissue composition measures in this study also matches very closely with results from our previous HM-MRI study with a smaller sample size (21 patients) by Chatterjee et al. [11]. In another study by Chatterjee et al. [17], the correlation between tissue measures from HM-MRI and quantitative histology using a semi-automatic software using Image Pro Premier (Media Cybernetics, Rockville, MD) on the basis of color, intensity, morphology, and background with the “Smart Segment” tool was performed on 25 patients. Strong correlation between measures for epithelium (CCC = 0.90, r = 0.93) and lumen volume (CCC = 0.87, r = 0.90) was reported. In this study as well we found similarly strong correlation between HM-MRI and pathologists’ measure of prostate tissue (epithelium: CCC = 0.73, r = 0.78 and lumen: CCC = 0.50, r = 0.72). In addition, the tissue composition measured in this study closely matches measures for prostate tissue from similar regions of interests in this previous study: epithelium volume (HM-MRI: 31% ± 15, quantitative histology: 34% ± 15), and lumen (HM-MRI: 24% ± 13, quantitative histology: 22% ± 11). This also suggests that tissue composition estimation using HM-MRI are reproducible. The current study used a larger dataset compared to the studies mentioned above. In addition, this study involved the consensus of 3 expert pathologists to derive the reference standard, which is better reference standard than any of these previously published studies. Therefore, in this larger statistically powered study involving 3 expert pathologists to derive the reference standard, our data reveals that the agreement in tissue composition measurement using hybrid multidimensional MRI and consensus of pathologists is on par with the inter-raters (pathologists) agreement.

Previous prostate microstructure imaging studies have demonstrated correlation of MRI-markers with histology. Luminal water imaging was performed in vivo in 17 patients and found to have similarly strong correlation (ρ = 0.75) between luminal water fraction and lumen fractional volume. However, luminal water imaging is not able to estimate epithelium volume. Vascular extracellular and restricted diffusion for cytometry in tumors (VERDICT) measurements of intracellular (r = 0.96), extracellular-extravascular (r = 0.96), and vascular (r = 0.39) volume fraction in three patients in vivo showed moderate to strong correlation [27]. Like HM-MRI, diffusion-relaxation correlation spectrum imaging [28] predicts volume fractions of stroma, lumen, and epithelium. However, it has been reported to have only moderate correlation with measurements of stroma (ρ = 0.32), epithelium (ρ = 0.80), and lumen (ρ = 0.57) on histology. Histological validation of diffusion-relaxation correlation spectrum imaging was also limited to nine ex vivo samples in this study and has not been evaluated in vivo [28].

The trend of cancers having increased epithelium and reduced lumen volume compared to benign tissue is seen here for both HM-MRI and pathologists’ measurement, which matches results in the previous studies mentioned above [11–13, 17, 26]. The ROC analysis showed excellent area under the curve of 0.768–0.969 for differentiating between cancer and benign tissue. This is very similar to those reported using tissue composition measured using HM-MRI and quantitative histology software (0.88–0.96). This diagnostic performance is similar or better than the reported performance achieved through the visual assessment of multiparametric MRI by radiologists using the PI-RADS guidelines [29, 30] and some of other quantitative MRI methods such as restriction spectrum imaging and luminal water imaging [31, 32]. This suggests that tissue composition measured using HM-MRI demonstrate not only excellent correlation with reference standard measures from histology by pathologists, but can also be used to diagnose prostate cancer. A prospective validation of HM-MRI to guide prostate biopsy and in diagnosing cancer in currently underway in a registered clinical trial (ClinicalTrials.gov Identifier: NCT03585660).

Our study has a few limitations. The tissue composition measured in vivo using HM-MRI was compared to pathologists’ assessment of formalin fixed prostate tissue ex vivo. Formalin fixation reduces the size of prostate compared to in vivo prostate (~ 15% reduction in volume) [33]. However, the effect of fixation on volume fractions of individual gland components has not been studied. Nonetheless, we expect these effects to be minimal. Benign features (such as inflammation, benign prostatic hyperplasia, inflammation, and prostatitis) that can mimic cancer [34, 35] were not specifically included in the analysis as separate groups. However, these pathologies were included as part of benign tissue from either of these 3 distinct prostatic zone regions to be better representative of the tissue. Another limitation of our model is that estimation of other histologic features (such as inflammation and blood vessels) that are not prominent features in prostate tissue (make up less than 5% typically) were not part of either MRI modeling or histologic analysis by pathologists. The validation was limited to MR images from Philips MR scanners (Achieva and Ingenia). Similar validation using different MR vendors is also needed in the future. The use of an endorectal potentially limits the generalization of the technique and needs to be tested without endorectal coil.

In conclusion, the results of tissue composition measurement from MRI compared to reference standard results from the consensus of 3 pathologists, reveal that ninety percent of absolute paired differences were within 18.8% and 22.4% in measuring epithelium and lumen, respectively. We are 95% confident that 90% of absolute paired differences were within 20.6% and 24.2% in measuring epithelium and lumen, respectively. These were less than our criterion of 30% and therefore we accept the agreement performance of HM-MRI. These absolute paired differences between HM-MRI and consensus results were less between pair of pathologists (inter-rater agreement). We also conclude that the between methods agreement is on par with the inter-raters (pathologists) agreement. In addition, high area under the receiver operating characteristic curve using tissue composition measures from HM-MRI and pathologists suggests that HM-MRI can potentially be used for non-invasive prostate cancer diagnosis with prostate cancers characterized by increased epithelium and reduced luminal volume compared to benign tissue.