Short-term repeatability of monoexponential apparent diffusion coefficient in prostate cancer: dependence on b-value,tumor grade,lesion size,and lesion location

Introduction

Diffusion-weighted imaging (DWI) is a cornerstone for the detection and characterization of prostate cancer (PCa). Prostate DWI signal is the most commonly fitted using the monoexponential function leading to monoexponential apparent diffusion coefficient (ADCm) values.

Although ADCm provides quantitative values (mm²/s), there are several factors affecting ADCm values, such as presence of noise. These factors may impact repeatability and reproducibility of ADCm values and different postprocessing methods. In routine clinical practice, ADCm parametric maps are usually only reviewed visually, and its quantitative properties are not commonly leveraged despite inherent quantitative nature.

Several studies have reported variations in ADCm values depending on the acquisition parameters, especially dependency on the used b-value (1,2), including in the PCa ADCm (3,4). In addition to DWI acquisition parameters, factors such as PCa lesion size (5) and PCa aggressiveness (6) have been observed to impact prostate DWI diagnostic performance and repeatability. These aspects are important to evaluate, including different noise characteristics, and to explore their impact on repeatability.

Repeatability of ADCm of PCa has been explored only in a few studies. In a study by Barrett et al. (7), 10 individuals were evaluated in a short test–retest set-up, same day repeated DWI. The results demonstrated good repeatability for 10th and 90th percentiles and median of ADC, based on Bland–Altman plots. In a previous study of 112 short-term repeated DWI scans (8), kurtosis apparent diffusion coefficient (ADCk) calculated from kurtosis function was evaluated in the context of various radiomics using machine-learning methods. These advanced radiomics of ADCk parametric maps demonstrated good repeatability and outperformed more commonly used region of interest (ROI)-based first order statistics. In a study by Rogers et al. (9), ADC from the VERDICT model was evaluated for repeatability in 41 men, evaluating five b-value combinations with b = 0 s/mm² and 500, 1500, and 2000, and 90, 500, finding that b-value set 0, 1500 was most repeatable. In context of radiotherapy planning, 11 patients were scanned in consecutive days, with highest limits of Bland–Altman agreement (−0.27 to 0.31 ×10⁻³ mm²/s) (10). Quantitative Imaging Biomarkers Alliance (QIBA) recommends making further repeatability studies for identification of quantitative imaging biomarkers (11). QIBA stresses the importance of assessing repeatability of imaging biomarkers with test–retest scans, and use of within-subject standard deviation (wSD) and coefficient of variation (wCV) for technical performance metrics (12).

In this study, the repeatability ADCm (calculated using DWI datasets acquired using 12 b-values [0–2000 s/mm²] and using different b-value combinations) was evaluated. The impact of different PCa lesion sizes and lesion location was explored. We aimed to provide evidence for technical performance of ADCm as a quantitative biomarker using QIBA proposed methodology.

Material and Methods

This study was approved by the institutional review board, and each patient gave written informed consent before enrolling in the study. We used imaging data, as described in Turku University Hospital, Turku, Finland (approval number 112/180/2012) (8), summarized below: all DWI examinations were performed at 3 T (Ingenuity PET/MR; Philips, Cleveland, OH, USA) with spin-echo sequence with single-shot echo-planar read-out, monopolar diffusion gradient scheme, three diffusion directions per each b-value, diffusion gradient timing (Δ) 24.5 ms, diffusion gradient duration (δ) 12.6 ms, diffusion time (Δ − δ/3) 20.3 ms, with TR/TE 3141/51 ms; field of view (FOV) 250 × 250 mm, acquisition matrix 100 × 99, reconstruction matrix 224 × 224, slice thickness 5 mm, and 12 b-values (number of signal averages) 0 (2), 100 (2), 300 (2), 500 (2) (2), 700 (2), 900 (2), 1100 (2), 1300 (2), 1500 (2), 1700 (3), 1900 (4), and 2000 (4) s/mm². The acquisition time was 8 min 29 s. Transversal T2-weighted images were acquired using a single-shot turbo spin-echo sequence with TR/TE 4668/130 ms, FOV 250 × 320 mm, acquisition matrix size 250 × 320, reconstruction matrix size 512 × 672, and slice thickness 2.5 mm. The DWI datasets were acquired as part of two prostate MRI examinations performed on the same day, shortly after each other. After the first MRI examination, the patient was taken out of the MRI room and asked to rest for 10–15 min. Subsequently, the second MR examination was performed after repositioning of the patient on the MR table.

PCa areas on DWI and T2-weighted images were manually delineated using whole mount prostatectomy sections by one radiologist with 10 years of experience, conducted individually for the first and second scans. Gleason Grade Group (GGG) values were assigned based on whole mount prostatectomy sections. A total of 170 ROIs were included, each with test and retest ADCm mean values, in 112 patients with PCa. No ROIs in the non-cancer-containing/benign tissue areas were delineated, and these areas were calculated as a whole gland mask minus PCa mask(s).

ADCm parameter maps with two b-value combinations from 0, 300 to 0, 2000 s/mm² were calculated with the commonly used formula where signal is normalized with b = 0 s/mm², as:

A D C m 0 , b = − ln ( S ( b ) / S ( 0 ) ) b

(1)

In addition, ADCm parameter maps with two b-value combinations excluding b-values of 0 s/mm², without signal normalization, were calculated. Mean ADCm values inside PCa lesions were evaluated using Spearman’s rank correlation and Bland–Altman plots. The latter were calculated to demonstrate differences between test and retest scans. In addition, two metrics recommended by QIBA were utilized: first, a proxy measure with two repeated measurements for within-subject standard deviation (denoted as wSD_proxy); and second, within-subject coefficient of variation (wCV) for validating the radiomics (11). These are defined as (11):

w S D p r o x y = [ ∑ N s u b j e c t s ( s c a n 1 − s c a n 2 ) 2 2 ] / N

(2)

w C V = ∑ i = 1 N ( s c a n 1 , i − s c a n 2 , i ) 2 2 N ( s c a n 1 , i + s c a n 2 , i 2 ) 2

(3)

The area under the receiver operating characteristic (ROC) curve (AUC) for classification potential between all GGGs 1–5 were calculated. In addition to GGG 1–5 in lesions, a simulated benign tissue, denoted as GGG = 0, was calculated from ADCm voxel level parametric maps of the whole prostate by excluding PCa lesions. Multiclass AUC (GGG = 1 to GGG = 5) and binary AUC (between GGG = 0, 1 vs higher) were calculated. Effect of b-value was evaluated for wSD_proxy equation (2) and wCV equation (3). The statistical calculations were done with R (version 4.2.3). P-values were reported as non-corrected unless otherwise noted, and p < 0.05 were considered as statistically significant.

Results

Fig. 1 shows Bland-Altman plots between the test and retest scans, with the best limits of agreement from −0.14 to 0.12 × 10⁻³ mm²/s, bias −0.01 × 10⁻³ mm²/s, with b = 0,2000 s/mm², and no observable trend differences across ADCm intensity range. Similar limits of agreement were observed when peripheral zone (PZ) and central zone (CZ) lesions were analyzed separately (Supplementary Information Figure S1a and S1b), and also when the first b-value for ADCm fit was selected to be 300 s/mm², meaning excluding b-values of 0, 100 s/mm² (Supplementary Information Figure S2a, S2b and S2c).

OPEN IN VIEWER

Fig. 1.

Bland–Altman plot for mean ADCm value difference between test and retest scans with 112 prostate cancer lesions, for ADCm parameter maps with two b-values from 0, 300 to 0, 2000 s/mm². ADCm, monoexponential apparent diffusion coefficient.

Differences in ADCm mean values between PCa lesions with different GGG (see Fig. 2, for PZ and CZ separately, Supplementary Information Figure S3a and S3b) reached the level statistical significance with b = 0,500 s/mm² and higher (Kruskal-Wallis p < 0.05). The same was observed also when the first b-value for ADCm fit was selected to be 300 s/mm², excluding b-values of 0 and 100 s/mm² (Supplementary Information Figure S4a, S4b, S4c). No significant reduction in wSD between GGG=0, and GGG≥1 was observed, however, the repeatability of ADCm mean in PCa lesions with GGG≥3 demonstrated a trend towards higher repeatability (Fig. 3). Similarly, wCV decreased towards higher GGG compared with GGG=1. Specifically, wCV was <0.07 with b-value sets from 0,700 to 0,1300 s/mm². Most of the evaluated lesions which had notable wSD up to 0.2 × 10⁻³ mm²/s were found in lower GGG.

OPEN IN VIEWER

Fig. 2.

Differences between GGGs 1–5 and benign tissue in 112 individuals for mean ADCm values inside prostate cancer region of interest for ADCm parameter maps with two b-values from 0, 300 to 0, 2000 s/mm²; the top is for the lesions in the peripheral zone while the bottom row is for the lesions in in the central gland. In total, 170 prostate cancer lesions per repetition were included. Note: GGG 0 represents benign tissue and is calculated as a whole gland mask minus prostate cancer mask(s). ADCm, monoexponential apparent diffusion coefficient; GGG, Gleason Grade Group.

OPEN IN VIEWER

Fig. 3.

Differences between GGGs 1–5 and benign tissue in 112 individuals for within-subject variance (wSD) and coefficient of variation (wCV) values in test–retest evaluations, for selected ADCm parameter maps with two b-values from 0, 700 to 0, 1300 s/mm². Note: GGG 0 represents benign tissue and is calculated as a whole gland mask minus prostate cancer mask(s). ADCm, monoexponential apparent diffusion coefficient; GGG, Gleason Grade Group.

When looking at the effect of different b-value acquisitions on ADCm performance (Fig. 4 and Supplementary Information Table S1), the multiclass AUC for characterization for GGG = 1 to GGG = 5 was gradually rising with the b-value, with biggest improvements before b = 500 s/mm², after which the improvement was found to be smaller. Similarly, wSD_, and wCV decreased. The Spearman correlation of ADCm with GGG 1–5 and corresponding p-value had the biggest improvements occurring up to b = 500 s/mm², after which differences were found to be negligible.

OPEN IN VIEWER

Fig. 4.

Area under the receiver operating characteristic curve (AUC), Spearman's correlation coefficient, within-subject variance (wSD), and within-subject coefficient of variance (wCV), for test–retest prostate lesion ADC calculated with two b-values, with maximum b-values in the range of 100–2000 s/mm². ADC, apparent diffusion coefficient.

We also plotted wSD against the PCa lesion; see the plots for b = 0, 700 to 0, 1300 s/mm² in Figure S5–S7 in the supplemental material. The lowest repeatability was observed with lesion size <5 mL while for lesion size 5 mL and higher, the wSD was <0.1 ×10⁻³ mm²/s. Dependence of lesion size is shown in Figure S8 in the supplemental material, where wSD was found to be under 0.1 ×10⁻³ mm²/s for lesions >3 mL (similarly when all 12 b-values were used and for PZ and CZ) (Figure S9 in the supplemental material).

Discussion

In this study, ADCm mean from PCa region was evaluated as a quantitative biomarker for PCa detection and characterization in 112 patients with PCa with short-term repeated DWI scans performed on the same day. The repeatability analysis for ADCm calculated using 2 b-values indicated that the biggest benefit from increasing b-value is gained with b-values higher than 500 s/mm², after which performance improvements are smaller. It is important to notice that the higher b-values require stronger motion probing gradients, leading to longer TE; however, in the present study, the impact of signal-to-noise ratio and TE was not evaluated. All b-values were acquired with the same TE. Based on our results and considering deviation of DWI signal decay from the monoexponential decay at higher b-values, in our datasets, we consider maximum reasonable b-values for monoexponential DWI signal quantification of PCa to be up 1000–1300 s/mm². In our experiments, we did not normalize the ADCm mean values across individuals, for more realistic technical evaluation as quantitative biomarker in each individual patient.

The monoexponential model provided good overall performance considering AUC, wSD, and wCV together. We focused particularly on mean ACDm, calculated from PCa lesions manually delineated using whole mount prostatectomy sections, upon request by the QIBA DWI committee to provide reference values for ADCm for the QIBA DWI statements regarding repeatability of ADCm.

The present study has some limitations, similar to our original manuscript using the same datasets (8), including the following: (i) no correction for possible correlation between multiple tumors in one PCa patients since 46/112 (41%) patients had more than one PCa lesion; (ii) no differentiation between benign tissues of peripheral zone or central gland. The benign tissue calculated as whole gland mask minus tumor masks. It is well established that ADCm values of the central gland, especially in the presence of benign prostatic hyperplasia, are lower than those of peripheral zone. This can at least partly explain our relative low classification performance for benign tissue versus PCa, with low AUC values; (iii) different b-value combinations were evaluated but it is important to stress that our results are influenced by the DWI acquisition parameters, such as TE (51 ms) and diffusion time (20.3 ms), and impact of these parameters was not evaluated. Moreover, the dependence of ADCm performance measures (both for PCa detection and characterization as well as repeatability) on signal-to-noise was not assessed; (iv) only the monoexponential model was applied (3,6); however, kurtosis function, which has higher information content (fitting quality), similar repeatability, and similar robustness against noise (13), was not evaluated. In addition, it is important to highlight that the acquisition parameters for DWI are not fully meeting the PI-RADS recommendations. Specifically, the slice thickness per PI-RADs should be ≤4 mm; however, in the present study we used a slice thickness of 5 mm (no gap) to ensure sufficient signal at high b-values. The remaining parameters are within the PI-RADS recommended parameter range (TE = ≤90 ms; TR = ≥3000 ms, FOV = 16–22 cm, in-plane dimension = ≤2.5 mm phase and frequency). The delay between scans was short (under 15 min), which is likely too short for physiological factors such as blood volume/hydration status, hemodynamic effects, and temperature to significantly vary. Lastly, the broader application of our results is limited by only a single MR scanner nature of the study, and we have no data from other MR scanners.

In conclusion, this technical note provides further evidence that ADC calculated with the monoexponential model (ADCm) can be a reliable quantitative imaging biomarker for PCa detection and characterization.

Supplemental Material