Classification of Prostate Tissue Using Descriptive Signal Enhancement Features Derived From DCE-MRI Acquisition With High Spatiotemporal Resolution

Abstract

Purpose:

The aim of this study was to investigate the diagnostic value of descriptive prostate perfusion parameters derived from signal enhancement curves acquired using golden-angle radial sparse parallel dynamic contrast-enhanced (DCE) magnetic resonance imaging (MRI) with high spatiotemporal resolution in advanced, quantitative evaluation of prostate cancer compared with the usage of apparent diffusion coefficient (ADC) values.

Methods:

A retrospective study (from January 2016 to July 2019) including 75 subjects (mean, 65 years; 46–80 years) with 2.5-second temporal resolution DCE-MRI and PIRADS 4 or 5 lesions was performed. Fifty-four subjects had biopsy-proven prostate cancer (Gleason 6, 15; Gleason 7, 20; Gleason 8, 13; Gleason 9, 6), whereas 21 subjects had negative MRI/ultrasound fusion-guided biopsies. Voxel-wise analysis of contrast signal enhancement was performed for all time points using custom-developed software, including automatic arterial input function detection. Seven descriptive parameter maps were calculated: normalized maximum signal intensity, time to start, time to maximum, time-to-maximum slope, and maximum slope with normalization on maximum signal and the arterial input function (SMN1, SMN2). The parameters were compared with ADC using multiparametric machine-learning models to determine classification accuracy. A Wilcoxon test was used for the hypothesis test and the Spearman coefficient for correlation.

Results:

There were significant differences (P < 0.05) for all 7 DCE-derived parameters between the normal peripheral zone versus PIRADS 4 or 5 lesions and the biopsy-positive versus biopsy-negative lesions. Multiparametric analysis showed better performance when combining ADC + DCE as input (accuracy/sensitivity/specificity, 97%/93%/100%) relative to ADC alone (accuracy/sensitivity/specificity, 94%/95%/95%) and to DCE alone (accuracy/sensitivity/specificity, 78%/79%/77%) in differentiating the normal peripheral zone from PIRADS lesions, biopsy-positive versus biopsy-negative lesions (accuracy/sensitivity/specificity, 68%/33%/81%), and Gleason 6 versus ≥7 prostate cancer (accuracy/sensitivity/specificity, 69%/60%/72%).

Conclusions:

Descriptive perfusion characteristics derived from high-resolution DCE-MRI using model-free computations show significant differences between normal and cancerous tissue but do not reach the accuracy achieved with solely ADC-based classification. Combining ADC with DCE-based input features improved classification accuracy for PIRADS lesions, discrimination of biopsy-positive versus biopsy-negative lesions, and differentiation between Gleason 6 versus Gleason ≥7 lesions.

Prostate cancer is the second most common cancer diagnosis in men and 1 of the leading causes of death^1,2 worldwide. Although randomized controlled trials showed a reduction of mortality up to 50% for screening programs,³ optimized analysis tools that make the screening procedure fast, reliable, and cost-efficient are needed.

Multiparametric magnetic resonance imaging (MRI) has proven to be a key component for the detection of clinically significant prostate cancer and for planning targeted biopsies as a subsequent step in the clinical pathway.^4,5 According to the current imaging guidelines published by the American College of Radiology, PIRADS v2.1 (Prostate Imaging and Reporting Archiving Data System), the acquisition of diffusion-weighted images (DWIs) and T2-weighted MR images is considered mandatory and most relevant for image-based evaluation of prostate cancer.⁶ Dynamic contrast-enhanced (DCE) MRI, which captures the tissue’s signal enhancement after injection of a gadolinium-based contrast agent, is also recommended to be included in the prostate examination protocols, especially as backup sequence, but it has lost relevance compared with the earlier versions of the PIRADS guidelines. Specifically, the current PIRADS v2.1 scoring system relies on DCE-MRI only for distinguishing between PIRADS 3 and 4 lesions in the peripheral zone,⁶ which has sparked a discussion on the usefulness of acquiring DCE-MRI as part of routine prostate examinations.

A major limitation of DCE-MRI is that there is no consensus on how the dynamic images should be analyzed. To cope with the large amount of images, many institutions use pharmacokinetic modeling approaches, often based on the classical Tofts or Kety models,^7,8 which have originally been proposed for 1 specific tissue type (eg, assessment of blood-brain barrier defects in the case of the Tofts model) and which have subsequently been applied across the body. Significant differences exist among commercial implementations of these modeling approaches, which result in high variability and low comparability between different software solutions, scanners, and users.^9,10 Therefore, the question arises if the downgrading of DCE-MRI in the PIRADS guidelines could possibly be related to inaccuracies introduced during data analysis because of the lack of a standardized and reproducible image-processing strategy.

In addition to the variability in data analysis, also the DCE-MRI acquisition techniques and protocols differ significantly across institutions. Novel DCE-MRI techniques have been presented over the last years that provide much higher spatial and temporal resolution compared with the techniques traditionally used for dynamic MR examination. Multiple studies have investigated the clinical use of these new techniques and have concluded that the higher spatiotemporal resolution translates into better diagnostic performance in discriminating normal prostate tissue and PIRADS 4 or 5 lesions.¹¹ This, again, leads to the question whether DCE-MRI could play a bigger role for the detection of prostate cancer if higher spatial and, especially, higher temporal resolution would be used routinely.

Therefore, the aim of this study is to reinvestigate the role of DCE-MRI for classifying PIRADS 4 or 5 prostate lesions when using a modern acquisition technique with high spatiotemporal resolution and when calculating descriptive enhancement features from the dynamic images, which have been designed to capture the observed enhancement characteristics of cancerous prostate lesions. Our hypothesis is that DCE-MRI–derived model-free enhancement features contribute to the detection of prostate cancer and improve the diagnostic performance when combined with apparent diffusion coefficient (ADC) and T2-weighted images.

MATERIALS AND METHODS

Study Population

Our institutional review board approved this study. All participants gave informed consent. Seventy-five male subjects (mean, 65 years; 46–80 years) who underwent multiparametric DCE prostate MRI at our institution from January 2016 to July 2019 with the diagnosis of PIRADS 4 or 5 lesions were retrospectively included (Fig. 1). Lesions were classified in consensus reading by 2 radiologists with subspecialty in abdominal imaging (D.T.B. with 16 years of experience, T.J.H. with 10 years of experience) using the PIRADS scoring system.⁶ All subjects underwent MRI/ultrasound fusion-guided targeted and random biopsies afterward. Fifty-four subjects had a positive biopsy with histopathology showing a prostate adenocarcinoma (Table 1). Twenty-one subjects had a negative targeted and random biopsy. In 13 subjects with a negative biopsy, histopathology revealed parenchymal inflammation. Seven subjects had a follow-up MRI (16.3 ± 4.5 months), which resulted in downgrading of the PIRADS 4 or 5 lesions to a lower PIRADS score.

FIGURE 1.

Overview of the included subjects and assignment to the groups that were analyzed.

TABLE 1.

Distribution of Age and Tumor Volume in the Different Subgroups

	n	Age ± SD, y	Tumor Volume ± SD, mL
Total	75	65 ± 7.8	1.1 ± 0.92
Negative biopsy	21	63 ± 7.1	0.34 ± 0.45*
Positive biopsy	54	65 ± 8.1	3.48 ± 0.58*
Gleason 6	15	64 ± 9.7	1.92 ± 0.80
Gleason 7	20	63 ± 6.9	0.58 ± 0.54
Gleason 8	13	68 ± 8.1	2.32 ± 1.80
Gleason 9	6	69 ± 5.9	0.79 ± 0.51
Gleason 7–9	39	65 ± 7.5	1.19 ± 1.11

^*Significant difference between biopsy-positive and biopsy-negative PIRADS lesions (P < 0.001).

Magnetic Resonance Imaging Examination

Magnetic resonance imaging examinations were performed using 3-T scanners (MAGNETOM Skyra; Siemens Healthcare GmbH, Erlangen, Germany). A routine prostate protocol was performed including an initial T2-weighted turbo spin echo sequence (repetition time [TR], 7.5 milliseconds; echo time [TE], 101 milliseconds; 3.5 mm ST; 0.56 × 0.56 mm; 320 × 320; acceleration factor 2) in 3 planes (axial, coronal, and sagittal), followed by a T1-weighted fat-saturated 3-dimensional GRE sequence precontrast (TR, 4.13 milliseconds; TE, 2.06 milliseconds; 3 mm ST; 0.82 × 1.2 mm; 288 × 202) and a RESOLVE DWI sequence (TR, 4.9 milliseconds; TE, 59 milliseconds; 3.675 mm ST; 1.69 × 1.69 mm; 118 × 118; b = 50/800; acceleration factor 2) afterward. Perfusion series were acquired last by using a golden-angle radial sparse parallel (GRASP) DCE-MRI sequence (TR, 4 milliseconds; TE, 2 milliseconds; 2.5 mm ST; 0.56 × 0.56 mm; 288 × 288) during the administration of gadoterate meglumine (0.01 mmol/kg, Dotarem; Guerbet, Villepinte, France). A 60-channel body coil was used (Body 60; Siemens Healthcare GmbH, Erlangen, Germany). Parallel imaging acceleration was achieved by using the generalized autocalibrating partial parallel acquisition. An injection delay of 20 seconds relative to the start of the GRASP sequence was used to ensure that sufficient precontrast data are acquired. The GRASP DCE-MRI technique combines golden-angle radial k-space sampling, compressed sensing, and parallel imaging reconstruction. Therefore, it can achieve a significantly higher spatial and temporal resolution compared with standard DCE techniques, which provide either less spatial or less temporal resolution.^12–15 Moreover, retrospective reconstruction with flexible temporal resolution is possible. Because of the small size of the lesions and the high vascularization of prostate tissue, high spatial as well as high temporal resolution are desired in prostate MRI examinations.^16,17 The capability of GRASP DCE-MRI to provide both simultaneously has been demonstrated in different organs and diseases.^18,19 For the current study, GRASP reconstructions with a temporal resolution of 2.5 seconds per time step were calculated. Reconstructions were performed on an external server using a custom-developed implementation of the GRASP algorithm and saved in the DICOM format for postprocessing analysis.

Magnetic Resonance Imaging/Ultrasound Fusion-Guided Biopsies

All biopsies were performed by board-certified urologists using a transrectal or transperineal approach. Transrectal MRI-targeted biopsies were performed using real-time virtual sonography (Preirus; Hitachi Medical Corporation, Tokyo, Japan) and a motion-tracking device (trakSTAR 3D Guidance). Transperineal MRI-targeted biopsies were performed with the iSR’obot Mona Lisa (Biobot, Singapore). In all subjects, a minimum of 3 targeted cores per lesion were taken. In addition, 12 to 18 random cluster biopsies were obtained in all cases.

Postprocessing of Dynamic Images

Because PIRADS 4 or 5 lesions are expected to show an earlier, more rapid, and more intense contrast enhancement compared with the normal peripheral zone (Fig. 2, top), the following descriptive parameters were calculated from the enhancement curves: the bolus arrival time (time to start [TTS]), defined as the time point before each curve reaches 10% of its overall maximum; the maximum slope (SM); the time to the maximum of the curve’s slope (TTSM); the time to maximum (TTM); and the normalized maximum signal intensity (Fig. 2, bottom). Image-based analysis was performed in MATLAB (2020b; MathWorks, Natick, MA) after loading and transforming the DCE-MRI DICOM images into a 4-dimensional array. Computations were performed on a standard laptop without parallel computing (M1, 2020: 8-core CPU, maximum 3.2 GHz, 16-GB RAM; Apple Inc, Cupertino, CA). Arterial input function (AIF) detection was implemented based on an approach described by Chen et al²⁰ for tumor assessment.^20,21 The bolus arrival time (TTS) was calculated for the AIF. Smoothing of the 4-dimensional array was performed using a Gaussian filter for the spatial dimensions and a median filter for the temporal dimension. Afterward, subtraction images were generated, and a “signal intensity over time” curve SI_v(t) was calculated for each voxel from the subtraction images.

FIGURE 2.

Top, Enhancement of an examplaric PIRADS 5 lesion (biopsy-confirmed Gleason 7 prostate cancer) compared with normal peripheral zone (PZ). Bottom, Illustration of the calculated descriptive perfusion parameters for the time-signal intensity curve of the lesion with the maximum slope (red dot), time to start (TTS), time-to-maximum slope (TTSM), and time to maximum (TTM) referenced to the bolus arrival time of the arterial input function (vertical yellow line).

In the next step, TTS, TTSM, TTM, SM, and the normalized maximum signal intensity (normalized to the signal of the 5% voxel with the highest signal per slice) were calculated. All temporal parameters were estimated relative to the bolus arrival time of the AIF, so that the values become independent of timing inaccuracies during the sequence acquisition. In addition, the SM was calculated for the time-signal intensity curves normalized to the maximum signal of each voxel (SMN1 = maximum slope normalized 1) and the maximum slope of the AIF (SMN2 = maximum slope normalized 2). Maps for these 7 values were generated in a pixel-by-pixel manner. To evaluate the influence of the temporal resolution, GRASP DCE images with a resolution of 5 seconds and 7.5 seconds (instead of 2.5 seconds) were calculated through temporal averaging, and maps were generated for the images with lower resolution as well.

The PIRADS lesions and the normal peripheral zone on the contralateral side were segmented on GRASP DCE images in accordance with the ADC maps (Fig. 3) by a radiologist with 3 years of experience in abdominal imaging (H.C.B.). This was done by marking the whole lesion on all slices (volumetric segmentation) and creating a 3-dimensional mask. Segmentations were done for the time point with optimal visual contrast of the lesion. A second reader (D.J.W.) performed segmentation of 10 randomly selected lesions out of the study cohort to assess interreader variability and the duration needed for segmentation. Using the created segmentation masks (Fig. 4), the mean, median, standard deviation, first and third quartile, skewness, kurtosis, and entropy inside the lesion were calculated from the TTS, TTM, TTSM, SM, SMN1, SMN2, and the normalized signal intensity maps.

FIGURE 3.

PIRADS 5 lesion (white arrow) with correlating lowered ADC (A) and low signal in T2-weighted images (B) in the left posteromedial to posterolateral zone showing an early (C, 51 seconds post injection) and more intense (D, 118 seconds post injection) contrast enhancement in the GRASP DCE series.

FIGURE 4.

GRASP DCE-derived maps of a PIRADS 5 lesion (white arrow) in the posteromedial to posterolateral left PZ showing a more rapid (maximum slope SM in SI/s, normalized maximum slope SMN in 1/s), more intense contrast (normalized signal intensity NSI), and earlier (TTS in seconds, TTSM in seconds, TTM in seconds) enhancement.

Statistical Data Analysis

Statistical analysis was done with commercially available software (JMP 14.0.0; SAS Institute Inc, Cary, NC). A Shapiro-Wilk test for normal distribution, Wilcoxon test for the hypothesis test, and the Spearman coefficient for correlation were calculated. Intraclass correlations were used to estimate the interreader variability.

Further data modeling was performed with MATLAB (2020b; MathWorks, Natick, MA). Three different machine-learning models were used for the evaluation of the multiparametric data: support vector machine (SVM) classification, which is a supervised method for classification and regression by hyperplanes and was introduced by Vapnik,^22,23 linear discriminant analysis (LDA), which uses linear combinations of features for classification and was introduced by Fisher,²⁴ and bootstrap-aggregated decision trees, which performs classification by creating a linear combination of multiple different decision trees.^23,25,26 Five-fold cross-validation was used for all models as a measure to evaluate the classification performance in small data sets.^27,28 The algorithms were applied on 3 binary classes: discrimination of PIRADS 4 or 5 lesion versus normal peripheral zone, discrimination of cancer-positive versus cancer-negative biopsies of PIRADS 4 or 5 lesions, and discrimination of Gleason 6 versus Gleason ≥7 in positive biopsies.

RESULTS

There were no adverse reactions or premature terminations of the examinations. The TTS map could not be calculated in 1 subject because of a low signal-to-noise ratio, which resulted in misdetection of the bolus arrival time in the tissue. This patient was excluded from the study. For all other subjects, calculation of all parameter maps worked without processing issues. Segmentation of the PIRADS lesion and the normal peripheral zone could be performed for all 75 subjects. The lesion volume (Table 1) was significantly lower in subjects with cancer-negative biopsy (0.34 ± 0.45 mL) than in subjects with cancer-positive biopsy (3.48 ± 0.58 mL, P = 0.0018). There was no significant difference between the tumor volume of Gleason 6 and Gleason ≥7 lesions. The computational time to perform the analysis was 65.7 ± 20.1 seconds. The average time for the segmentation of a lesion was 21.9 ± 5.3 seconds for reader A (H.C.B.) and 23.7 ± 6.7 seconds for reader B (D.J.W.). Intraclass correlation values ranged from 0.964 (SMN1) to 0.997 (TTS) for the interreader variability.

Evaluation of Descriptive Dynamic Contrast-Enhanced Features

The TTS, TTSM, and TTM were significantly lower, whereas SM, SMN1, SMN2, and the normalized maximum intensity were significantly higher for the PIRADS 4 or 5 lesions compared with the peripheral zone (P < 0.001, Table 2A), which reflects an earlier and more rapid enhancement of lesions compared with the normal peripheral zone (Fig. 5). All maximum slope values (SM: P = 0.03, SMN1: P = 0.01, SMN2: P = 0.03) were significantly higher, whereas TTSM (P = 0.01) and TTM (P = 0.01) were significantly lower for cancer-positive biopsies relative to cancer-negative biopsies (Table 2). The TTSM was significantly higher for Gleason 6 lesions than for Gleason 9 (P = 0.04). There were no significant differences between the Gleason scores for all other values and no significant differences between lesions with Gleason 6 score and those with Gleason ≥7. There was a significant negative moderate correlation between TTM and the Gleason score (P = 0.048, ρ = −0.27). There was a significant moderate negative correlation between SM (r = 0.31, P = 0.0076), TTSM (r = −0.38, P = 0.0008), as well as TTM (r = −0.3, P = 0.0112) and the Gleason score. SMN1 (r = 0.26, P = 0.0235), SMN2 (r = 0.28, P = 0.0165), TTS (r = −0.25, P = 0.029), and the normalized signal intensity (r = 0.26, P = 0.0228) showed a weak significant correlation with the Gleason score.

TABLE 2.

Overview of the Mean Value of the 7 DCE-Generated Parameters

A	n (PIRADS 4/5)	SM, SI/s	SMN1, 1/s	SMN2, 1/s	TTS, s	TTSM, s	TTM, s	Normed SI	ADC, mm2/s
Peripheral zone*	74	20.1 ± 12.7	0.1 ± 0.04	0.44 ± 0.18	13.8 ± 4.9	36 ± 32	218.3 ± 49	0.9 ± 0.24	1821 ± 248
Lesion*	74 (33/41)	38.6 ± 18.9	0.15 ± 0.05	0.69 ± 0.21	11.5 ± 5.7	23.6 ± 16	154 ± 68.3	1.1 ± 0.19	1005 ± 265
False-positive†	20 (11/9)	30.9 ± 18.8	0.13 ± 0.05	0.58 ± 0.17	12.4 ± 6.3	26.5 ± 16.6	182 ± 64	1.04 ± 0.24	1130 ± 271
True-positive†	54 (22/32)	41.6 ± 17.5	0.17 ± 0.05	0.73 ± 0.27	11 ± 4	22.5 ± 14.3	143.2 ± 72.5	1.13 ± 0.24	956 ± 238
Gleason 6‡	15 (9/6)	38.7 ± 21.6	0.17 ± 0.06	0.68 ± 0.26	15.0 ± 9.9	26.7 ± 18.8	164.1 ± 61	1.09 ± 0.32	1034 ± 228
Gleason 7	20 (9/11)	37.5 ± 11.2	0.16 ± 0.04	0.74 ± 0.13	9.3 ± 3.8	21.3 ± 16	145.1 ± 66.3	1.1 ± 0.18	968 ± 313
Gleason 8	13 (2/11)	45.6 ± 19	0.17 ± 0.03	0.75 ± 0.09	10.3 ± 2.8	22.7 ± 19	115.2 ± 41.5	1.13 ± 0.21	889 ± 215
Gleason 9‡	6 (1/5)	53.6 ± 28.1	0.18 ± 0.05	0.81 ± 0.09	9.3 ± 2.5	15.4 ± 3.1	145.5 ± 83.3	1.33 ± 0.31	866 ± 286
Gleason 7–9	39 (12/27)	42.4 ± 17.8	0.17 ± 0.04	0.75 ± 0.11	9.8 ± 3.2	20.9 ± 15.8	61.5 ± 32.3	1.15 ± 0.21	932 ± 276
*		<0.001	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001
†		0.029	0.01	0.03	—	0.01	0.01		0.04
‡						0.04

B	SM, SI/s	SMN1, 1/s	SMN2, 1/s	TTS, s	TTSM, s	TTM, s	Normed SI
Peripheral zone*	15.94 ± 9.2	0.09 ± 0.03	0.48 ± 0.14	14.45 ± 5.35	37.1 ± 18.53	223.65 ± 47.3	0.85 ± 0.17
Lesion*	29.62 ± 14.04	0.12 ± 0.04	0.78 ± 0.14	11.75 ± 5.2	31.43 ± 7.7	160.2 ± 68.8	1.12 ± 0.2
False-positive	26.5 ± 13.92	0.11 ± 0.04	0.72 ± 0.17	13.25 ± 4.5	29.68 ± 5.63	182.35 ± 73.25	1 ± 0.14
True-positive	30.24 ± 14.53	0.17 ± 0.03	0.81 ± 0.12	11.2 ± 5.4	32.93 ± 9.93	152.6 ± 65.5	1.17 ± 0.19
Gleason 6	28.55 ± 16.23	0.12 ± 0.04	0.79 ± 0.17	12.55 ± 6.3	34.75 ± 11.2	168.25 ± 66.2	1.14 ± 0.21
Gleason 7	26.33 ± 10.7	0.12 ± 0.04	0.81 ± 0.11	9.4 ± 5.7	33.3 ± 11.2	152.4 ± 67.35	1.16 ± 0.17
Gleason 8	33.89 ± 14.22	0.13 ± 0.03	0.8 ± 0.06	12.2 ± 4.1	32.08 ± 11.2	134.6 ± 54.95	1.18 ± 0.18
Gleason 9‡	39.58 ± 18.86	0.14 ± 0.03	0.83 ± 0.09	11.05 ± 3.4	28.73 ± 11.93	150.15 ± 84	1.26 ± 0.26
Gleason 7–9	30.93 ± 13.97	0.13 ± 0.06	0.81 ± 0.09	10.6 ± 4.95	32.15 ± 6.73	146.25 ± 65.05	1.18 ± 0.18
*	<0.001	<0.001	<0.001	<0.004	<0.0063	<0.001	<0.001

C	SM, SI/s	SMN1, 1/s	SMN2, 1/s	TTS, s	TTSM, s	TTM, s	Normed SI
Peripheral zone*	12.49 ± 9.2	0.09 ± 0.03	0.49 ± 0.18	14.85 ± 5.7	54 ± 26.55	221.48 ± 48.38	0.86 ± 0.17
Lesion*	23.19 ± 14.04	0.12 ± 0.03	0.79 ± 0.14	12.15 ± 5.93	43.5 ± 12.38	160.13 ± 69.6	1.13 ± 0.2
False-positive	19.46 ± 13.92	0.11 ± 0.04	0.71 ± 0.18	15 ± 4.73	44.25 ± 19.5	185.25 ± 72.75	1 ± 0.2
True-positive	24.66 ± 14.53	0.13 ± 0.04	0.82 ± 0.12	11.1 ± 6	43.5 ± 8.25	150 ± 66.75	1.17 ± 0.13
Gleason 6	22.72 ± 16.23	0.13 ± 0.03	0.81 ± 0.09	10.8 ± 5.4	42 ± 6.38	144 ± 66	1.18 ± 0.19
Gleason 7	23.3 ± 10.7	0.12 ± 0.04	0.82 ± 0.1	9.9 ± 6	42.75 ± 6	151.5 ± 69	1.18 ± 0.2
Gleason 8	26.58 ± 14.22	0.13 ± 0.03	0.81 ± 0.08	11.78 ± 5.55	42.75 ± 6.75	131.25 ± 56.25	1.22 ± 0.24
Gleason 9‡	29.12 ± 18.86	0.13 ± 0.03	0.82 ± 0.07	11.18 ± 4.88	39.75 ± 6.75	150.75 ± 63.75	1.13 ± 0.22
Gleason 7–9	25.39 ± 13.97	0.14 ± 0.05	0.83 ± 0.19	11.85 ± 7.58	46.5 ± 12.15	166.5 ± 68.25	1.15 ± 0.21
*	<0.001	<0.001	<0.001	0.0024	0.0043	0.021	<0.001

Overview of the mean value of the 7 DCE-generated parameters with a temporal resolution of 2.5 seconds (A) and 5 seconds (B) (SM, SMN1/2, TTS, TTSM, TTM, SI) and the ADC for each group and subgroup with comparison of the PIRADS lesion and the tumor-free predicted contralateral peripheral zone, biopsy-positive and biopsy-negative lesions, and in dependence on the Gleason score.

^*,†,‡Significant differences with correlating P values.

DCE, dynamic contrast-enhanced; SM, maximum slope; SMN1, maximum slope normalized 1; SMN2, maximum slope normalized 2; TTS, time to start; TTSM, time to the maximum of the curve’s slope; TTM, time to maximum; ADC, apparent diffusion coefficient.

FIGURE 5.

ADC (apparent diffusion coefficient), TTM, and SMN1 (normalized maximum slope 1) in PIRADS lesions and normal PZ as well as in dependence on the Gleason score. ADC was significant (P < 0.05) higher, TTM was significant higher, and SMN1 was significant lower in the PZ compared with PIRADS 4 or 5 lesions. There was a significant correlation between Gleason score and ADC (weak negative), TTM (moderate negative), and SMN1 (weak positive).

At the reduced temporal resolution of 5 seconds and 7.5 seconds, TTS, TTSM, and TTM were also significantly lower, whereas maximum slope parameters (SM, SMN1, SMN2) were significant higher for a PIRADS lesion compared with the peripheral zone (Tables 2B, C). In contrast to DCE parameters derived from the temporal resolution of 2.5 seconds, there was no significant difference between cancer-positive and cancer-negative biopsies and no significant difference between different Gleason scores. There was a significant weak negative correlation between TTM and the Gleason score (r = −0.24, P = 0.045) and a significant moderate correlation between the normalized signal intensity and the Gleason score (r = 0.35, P = 0.022) at a temporal resolution of 5 seconds. Regarding the DCE values derived at a temporal resolution of 7.5 seconds, there was a significant negative, moderate correlation between Gleason and TTS (r = −0.31, P = 0.007), a significant negative, weak correlation between Gleason and TTSM (r = −025, P = 0.03), and a significant moderate correlation between the normalized signal intensity and the Gleason score (r = 0.39, P = 0.0008).

Evaluation of Apparent Diffusion Coefficient Values

The ADC values were significantly lower (1005 10⁻⁶ ± 265 mm²/s) for PIRADS lesions compared with the normal peripheral zone (1821 10⁻⁶ ± 248 mm²/s; P < 0.0001). The ADC values were significant lower for biopsy-positive PIRADS lesions (885 ± 260 10⁻⁶ mm²/s) than for biopsy-negative lesions (1020 ± 242 10⁻⁶ mm²/s). There were no significant ADC differences between Gleason 6 and 9 and no significant differences between Gleason 6 compared with 7 to 9. There was a significant weak negative correlation between ADC and the Gleason score (r = −0.27, P = 0.021; Fig. 5).

Multiparametric Analysis

Accuracy for the discrimination of the PIRADS lesion from normal peripheral zone was best by using DCE + ADC parameters for all 3 models with an accuracy of 97% compared with 94% for ADC alone. The discrimination of PIRADS lesions decreased using DCE parameters without ADC for all 3 models with an accuracy of 81.7% to 82.4% (Table 3A). Multiparametric analysis using ADC + DCE values showed the best discrimination of biopsy-positive and biopsy-negative PIRADS 4 or 5 lesions with an accuracy of up to 82% (45% sensitivity, 99% specificity), whereas the discrimination using ADC provided lower accuracy, sensitivity, and specificity for all models (Table 4A). The best diagnostic performance regarding accuracy, sensitivity, and specificity for the differentiation of Gleason 6 and Gleason 7 to 9 cancer in biopsy-positive lesions was achieved with the SVM model for DCE values solely (76% accuracy, 43% sensitivity, 92% specificity; Table 4A).

TABLE 3.

Performance of the 3 Models (SVM, LDA, and Bagged Trees) for the Differentiation of PIRADS Lesions Compared With Normal Peripheral Zone by Using DCE + ADC Values and ADC or DCE Values Alone

A	DCE + ADC		DCE		ADC
	Accuracy (AUC)	Sensitivity/Specificity	Accuracy (AUC)	Sensitivity/Specificity	Accuracy (AUC)	Sensitivity/Specificity
SVM	0.97 (0.99)	0.93/1	0.824 (0.87)	0.83/0.82	0.93 (0.98)	0.96/0.91
LDA	0.97 (1)	0.96/0.97	0.817 (0.86)	0.83/0.8	0.94 (0.99)	0.95/0.95
Bagged trees	0.97 (0.99)	0.97/0.96	0.817 (0.85)	0.8/0.83	0.94 (0.99)	0.93/0.95

B	DCE + ADC		DCE
	Accuracy (AUC)	Sensitivity/Specificity	Accuracy (AUC)	Sensitivity/Specificity
SVM	0.937 (0.99)	0.96/0.92	0.74 (0.82)	0.8/0.68
LDA	0.93 (0.98)	0.94/0.92	0.72 (0.78)	0.79/0.75
Bagged trees	0.923 (0.98)	0.90/0.94	0.79 (0.86)	0.76/0.82

C	DCE + ADC		DCE
	Accuracy (AUC)	Sensitivity/Specificity	Accuracy (AUC)	Sensitivity/Specificity
SVM	0.915 (0.97)	0.93/0.9	0.74 (0.82)	0.8/0.68
LDA	0.92 (0.98)	0.93/0.92	0.77 (0.8)	0.8/0.73
Bagged trees	0.923 (0.98)	0.90/0.93	0.78 (0.86)	0.75/0.82

Dynamic contrast enhancement parameters were calculated for a temporal resolution of 2.5 seconds (A), 5 seconds (B), and 7.5 seconds (C).

SVM, support vector machine; LDA, linear discriminant analysis; DCE, dynamic contrast-enhanced; ADC, apparent diffusion coeffient.

TABLE 4.

Performance of the 3 Models (SVM, LDA, and Bagged Trees) for the Differentiation of PIRADS Lesions With Positive or Negative Biopsy for Prostate Cancer and for the Discrimination Between Gleason 6 and Gleason 7–9 Lesions

A		DCE + ADC		DCE		ADC
		Accuracy (AUC)	Sensitivity/Specificity	Accuracy (AUC)	Sensitivity/Specificity	Accuracy (AUC)	Sensitivity/Specificity
Biopsy positive/negative	SVM	0.82 (0.77)	0.45/0.99	0.8 (0.77)	0.4/0.96	0.72 (0.52)	0/1
	LDA	0.78 (0.77)	0.50/0.90	0.75 (0.76)	0.45/0.86	0.71 (0.6)	0.09/0.94
	Bagged trees	0.65 (0.6)	0.2/0.18	0.63 (0.67)	0.25/0.78	0.6 (0.48)	0.29/0.72
Gleason 6/7+	SVM	0.75 (0.76)	0.5/0.84	0.76 (0.66)	0.43/0.92	0.72 (0.58)	0/1
	LDA	0.69 (0.69)	0.6/0.72	0.63 (0.59)	0.4/0.72	0.67 (0.66)	0.07/0.92
	Bagged trees	0.72 (0.61)	0.33/0.87	0.72 (0.55)	0.33/0.87	0.63 (0.58)	0.33/0.74

B		DCE + ADC		DCE
		Accuracy (AUC)	Sensitivity/Specificity	Accuracy (AUC)	Sensitivity/Specificity
Biopsy positive/negative	SVM	0.71 (0.62)	0.24/0.8	0.6 (0.55)	0.25/0.84
	LDA	0.73 (0.76)	0.4/0.6	0.62 (0.57)	0.45/0.69
	Bagged trees	0.63 (0.6)	0.2/0.8	0.68 (0.55)	0.2/0.86
Gleason 6/7+	SVM	0.7 (0.54)	0.33/0.68	0.71 (0.53)	0.4/0.82
	LDA	0.7 (0.55)	0.27/0.89	0.68 (0.50)	0.2/0.89
	Bagged trees	0.72 (0.62)	0.47/0.83	0.66 (0.62)	0.27/0.83

C		DCE + ADC		DCE
		Accuracy (AUC)	Sensitivity/Specificity	Accuracy (AUC)	Sensitivity/Specificity
Biopsy positive/negative	SVM	0.72 (0.63)	1/0	0.59 (0.54)	0.2/0.83
	LDA	0.68 (0.75)	0.45/0.76	0.60 (0.55)	0.45/0.67
	Bagged trees	0.65 (0.61)	0.35/0.75	0.68 (0.55)	0.2/0.86
Gleason 6/7+	SVM	0.7 (0.53)	0.16/0.64	0.7 (0.50)	0.22/0.5
	LDA	0.65 (0.53)	0.19/0.79	0.63 (0.50)	0.19/0.86
	Bagged trees	0.68 (0.55)	0.11/0.86	0.66 (0.58)	0.16/0.79

Dynamic contrast enhancement parameters were calculated for a temporal resolution of 2.5 seconds (A), 5 seconds (B), and 7.5 seconds (C).

SVM, support vector machine; LDA, linear discriminant analysis; DCE, dynamic contrast-enhanced; ADC, apparent diffusion coeffient.

Multiparametric analysis for the discrimination of PIRADS lesions from normal peripheral zone was slightly worse by using DCE + ADC parameters with perfusion parameters derived from a DCE GRASP with lower temporal resolution with an accuracy of 93.7% at 5 seconds (Table 3B) and 92.3% at 2.5 seconds (Table 3C). Diagnostic accuracy by using DCE values generated with a temporal resolution of 5 seconds and 7.5 seconds was lower for the discrimination of biopsy-positive and biopsy-negative lesions, as well as for the discrimination of Gleason 6 and Gleason 7–9 cancer by using DCE parameters in combination with and without ADC values (Tables 4B, C).

DISCUSSION

The aim of our study was to establish robust, model-free parameters on the basis of GRASP DCE-MRI with its unique high spatiotemporal resolution, and to use these parameters for reassessing the value of DCE-MRI for prostate cancer detection. We generated 7 descriptive parameters with a low interreader variability from the voxel-wise analysis of signal enhancement curves. All parameters showed significant differences between the normal peripheral zone and PIRADS lesions, describing an earlier, more rapid, and increased enhancement in comparison to the normal peripheral zone. In addition, there was a significant correlation between the time point as well as the speed of contrast uptake and the Gleason score. Therefore, we conclude that DCE-MRI provides complementary information for the differentiation of the normal peripheral zone and prostate cancer.

Multiparametric MRI is a common and well-established method for the detection of prostate cancer⁵ and is useful to improve the quality of targeted prostate biopsies.^29,30 Several studies have shown higher accuracy for MRI prostate protocols that include DCE imaging,^31,32 whereas other studies showed no additional value.^33,34 This discrepancy may result from differences in the temporal and spatial resolution of the underlying imaging techniques on the one hand, and from the use of different pharmacokinetic models on the other hand, as de Rooij et al⁵ discussed in their multicentric meta-analysis. Recent studies have demonstrated the potential of GRASP DCE-MRI for the detection of cancer because of its uniquely high spatiotemporal resolution.^11,12

However, even when using an imaging technique with high spatiotemporal resolution, a residual problem consists in the low reproducibility of pharmacokinetic models, resulting in variabilities of up to 74.1% as demonstrated by Heye et al.^9,10 Therefore, model-free descriptive parameters may provide a more stable and more viable approach to evaluate the additional diagnostic value provided by DCE-MRI. The capability of model-free parameters has already been observed in a study by Rosenkrantz et al,³⁵ which obtained similar diagnostic value for the interpretation of maximum slope maps in comparison to model-derived parameters with a temporal resolution of 5.5 seconds. Here, we investigated 7 candidates for model-free parameters, which were derived voxel-by-voxel at a temporal resolution of 2.5, 5, and 7.5 seconds. Our results showed significantly earlier, more rapid, and increased enhancement of lesions compared with the normal peripheral zone. Moreover, all perfusion-derived parameters showed a significant correlation with the Gleason score at a temporal resolution of 2.5 seconds. The values derived from DCE GRASP with a temporal resolution of 5 seconds and 7.5 seconds showed a weaker or no significant correlation with the Gleason score. The results for these parameters correlate well with the visual observation of early contrast enhancement seen in those lesions^12,36 and are in line with other studies. For example, Zelhof et al³⁷ demonstrated in a quantitative analysis earlier and faster enhancement of malignant lesions compared with the peripheral zone based on variables that were derived from signal enhancement curves with temporal resolution of 7.78 to 8.36 seconds.

When combining these parameters additionally with ADC values, our study showed that higher accuracy is achieved in the quantitative assessment and discrimination of PIRADS 4 and 5 lesions versus the normal peripheral zone (based on PIRADS scoring system and biopsy). Moreover, our study indicated the potential value of the perfusion parameters and their intrinsic distribution in lesions as features to differentiate between prostate cancer with Gleason 6 and higher Gleason scores, which is relevant for the treatment of subjects and therapy outcome.³⁸ The diagnostic accuracy of 82% based on the combination of perfusion parameters and the ADC value with a sensitivity of 45% and a specificity of 99% agrees with the results of other studies that used machine learning approaches. For instance, Antonelli et al³⁹ obtained a sensitivity of 50% and a specificity of 88% for the discrimination of Gleason 6 and Gleason >6 cancer in 164 subjects by using ADC and 1 DCE-MRI feature (maximum enhancement). Current state-of-the-art imaging can still improve in the detection of suspicious lesions, and negative biopsies occur for PIRADS 5 lesions.^30,40 Especially the combination of ADC values and DCE showed potential for further differentiation of PIRADS 4 or 5 lesions regarding biopsy-negative lesions and the Gleason score in a multiparametric analysis. The diagnostic performance of the multiparametric models was worse by using DCE values derived from a temporal resolution of 5 seconds and 7.5 seconds. Recent studies demonstrated lower diagnostic accuracy for Tofts models as potential input layer for convolutional networks in comparison to DCE time series⁴¹ with the disadvantage of larger data amount as model input, which creates high computational cost. Therefore, the model-free parameter maps used in our study may serve as alternative for the input layer of convolutional networks in future studies. Taken altogether, DCE-derived parameters may provide a considerable contribution to the diagnostic performance of prostate MRI, especially because the computational costs are rather low and the analysis including segmentation could be performed in a short time if properly integrated into the reading software. Dynamic contrast-enhanced–derived parameters from a temporal resolution of 5 seconds (instead of 2.5 seconds) still showed similar performance when using multiparametric models, but worse performance in a monoparametric analysis. Therefore, DCE imaging with high spatiotemporal resolution, as achieved in this study with the GRASP DCE-MRI technique, seems to be the key for further improving prostate cancer detection and characterization. Further studies are needed to analyze the optimal temporal resolution and its importance for the diagnostic performance. Considering the controversial discussion about the administration of gadolinium-based contrast agents,^42,43 the acquisition and analysis of perfusion imaging might be adequate, especially for differentiation between Gleason 6 and Gleason 7 to 9 cancer because of the potential relevance for therapeutic management. Although more advanced ADC measurements could not provide additional information for the prediction of clinically significant prostate cancer,⁴⁴ acquisition of DCE imaging might be justified in these cases depending on further clinical risk factors. In addition, perfusion-derived parameters provide a viable backup in the case of missing or failed acquisition of DWI, for example, because of artifacts from hip replacements.

Our study has several limitations. First, it is a retrospective study with a small patient cohort. The cohort was established by selecting subjects with PIRADS 4 or 5 lesions for the development of the methodology and technical implementation of our analysis approach. Thus, the small sample size may introduce a selection bias. Nevertheless, we decided to validate our new method in a customized data set to evaluate the potential of model-free perfusion parameters and used k-fold cross-validation to use machine learning algorithms in the relatively small data set.⁴⁵ Twenty-one of 75 subjects with PIRADS 4 or 5 lesions had a histological negative biopsy with a significant lower lesion volume than subjects with biopsy-positive lesions, which results in potential sampling error and misinterpretation of lesions. A further explanation of biopsy-negative PIRDS 4 or 5 lesions is acute inflammation, which has been found in 13 of 21 negative biopsies and is a known confounder in prostate MRI imaging.⁴⁶ Nevertheless, because of the combination of MRI/ultrasound fusion-guided targeted and clustered biopsy, this approach seems to be reasonable and represents the clinical routine and best-available reference standard. Further evaluation in multicenter studies is needed, especially because another recent multicenter study demonstrated no improvement in discrimination of prostate cancer and normal tissue by adding quantitative DCE-derived parameters.⁴⁷ A different approach to yield more comparable and standardized data consists in the implementation of automated imaging workflows.⁴⁸

We demonstrated the potential of model-free DCE-derived feature parameters in addition to ADC values for the assessment of the peripheral zone. In a next step, the machine learning models need to be trained on larger data sets, including healthy subjects and a broader variety of lesions to create viable predictor models. Moreover, the use of clinical data and laboratory parameters (such as PSA level, age, or comorbidities) as input variables in combination with further promising image-derived parameters such as T2 mapping⁴⁹ or improved DWI sequences⁵⁰ might advance the prediction of clinically significant prostate cancer. In summary, additional studies are needed to evaluate and validate our method, as well as the general potential of high spatiotemporal DCE-MRI to improve the detection and characterization of prostate cancer.

CONCLUSIONS

We have developed a robust and intuitive method to assess GRASP DCE-MRI by voxel-wise analyses of the time-signal intensity curves. All perfusion parameters derived from the signal-time curves showed significant differences between the normal peripheral zone and PIRADS lesions. Multiparametric analysis combining perfusion-derived parameters and ADC showed improved diagnostic performance for the discrimination of PIRADS 4 or 5 lesions and the normal peripheral zone compared with ADC alone. Further classification of the multiparametric data with different machine learning algorithms showed promising but still improvable results for the evaluation of histopathological features such as the Gleason score or the differentiation of biopsy-negative but radiologically highly suspicious lesions.