Application of non-Gaussian water diffusional kurtosis imaging in the assessment of uterine tumors: A preliminary study

Objectives To evaluate the interobserver reliability and value of diffusional kurtosis imaging (DKI) in the assessment of uterine tumors compared with those of conventional diffusion-weighted imaging (DWI). Methods This retrospective study was approved by our institutional review board, which waived the requirement for informed consent. Fifty-eight women (mean age: 55.0 ± 13.6 years; range: 30–89 years) with suspected malignant uterine tumors underwent 3-T magnetic resonance imaging using DKI and DWI. Twelve had coexisting leiomyoma. Two observers analyzed region-of-interest measurements of diffusivity (D), kurtosis (K), and the apparent diffusion coefficient (ADC) of uterine lesions and healthy adjacent tissues. Interobserver agreement was evaluated using the intra-class correlation coefficient (ICC). The mean values were compared using one-way analysis of variance with a post-hoc Tukey’s honestly significant difference test. The diagnostic accuracy of D and ADC in differentiating malignant tumors from benign leiomyomas was analyzed using receiver operating characteristic (ROC) analysis. Results The ICCs between the two observers in evaluating D, K, and the ADC of the malignant tumors were higher than 0.84, suggesting excellent interobserver agreements. The mean D (×10−3 mm2/s) of uterine cancers (1.05 ± 0.41 and 1.09 ± 0.40 for observers 1 and 2, respectively) were significantly lower than those of leiomyoma (1.40 ± 0.37 and 1.56 ± 0.33, respectively; P < 0.05), healthy myometrium (1.72 ± 0.27 and 1.69 ± 0.30, respectively; P < 0.001), and healthy endometrium (1.53 ± 0.35 and 1.42 ± 0.37, respectively; P < 0.005). There was no significant difference in the area under the ROC curve between D and ADC. The mean K of uterine cancers (0.88 ± 0.28 and 0.90 ± 0.23, respectively) were higher than those of myometrium (0.72 ± 0.10 and 0.73 ± 0.10, respectively; P < 0.001), healthy endometrium (0.65 ± 0.13 and 0.60 ± 0.18, respectively; P < 0.001), and leiomyoma (0.76 ± 0.14 and 0.77 ± 0.16, respectively; not significant, P > 0.1). Conclusions Interobserver agreements in evaluating D, K, and ADC were moderate to excellent. D performed equally to conventional DWI in differentiating between benign and malignant uterine lesions. The mean K of malignant uterine lesions was significantly higher than that of non-tumorous myometrium or endometrium.


Introduction
Uterine cancers include cervical cancers and corpus cancers. Cervical cancer is the second most common cancer in women worldwide with nearly 530,000 new cases and 275,000 deaths attributed to the disease annually [1]. Corpus cancers, also referred to as endometrial cancers, are the most commonly diagnosed gynecologic malignancy among women in the US; an estimated 60,050 new cases and 10,470 deaths are expected in 2016 [2].
Over the last decade, several studies on functional magnetic resonance (MR) imaging have demonstrated the utility of functional MR imaging sequences, such as diffusion-weighted imaging (DWI), in the diagnosis, and pre-and post-operative assessment of uterine cancers [3][4][5][6][7][8][9][10]. Tumors are frequently more cellular than the tissue from which they originate, thus appearing to have a relatively high-signal intensity on DWI [3]. Conventional DWI studies the motion of water molecules assumed to undergo Gaussian diffusion. However, water in biologic tissues is restricted by its interactions with other molecules and cell membranes; therefore, the assumption of Gaussian water diffusion may be inadequate to describe the actual diffusion in tissues [11].
Diffusional kurtosis imaging (DKI) is a new and promising diffusion imaging technique that extends DWI through the quantification of non-Gaussian water diffusion and requires the use of multiple and higher b values. Recently, DKI was developed with the aim of characterizing the diffusional heterogeneity emerging from multiple tissue compartments with different diffusivities [11,12]. DKI provides novel in vivo diffusion properties that describe tissue microstructure by analyzing not only diffusivity (D) but also kurtosis (K), which is a unit-less index of non-Gaussianity [13,14]. DKI has been widely applied in recent years because of its clinical utility and robust theoretical framework [13,15]. To the best of our knowledge, however, the utility of DKI for the evaluation of uterine tumors has not yet been investigated. The purpose of this study was to evaluate the interobserver reliability and the value of DKI in the assessment of uterine tumors compared with that of conventional DWI. before therapy between November 15th, 2013 and February 4th, 2015. Fifty-eight women (mean age 55.0 ± 13.6 years [standard deviation], range 30-89 years) with suspected malignant uterine tumors were included in the study. After surgery (N = 26) or biopsy (N = 32), 53 had histopathologic proof of malignancy and the remaining five had no histopathologic proof of malignancy. For five of the 53 women who had histopathologically proven malignant tumors, it was difficult to measure signal intensity of the tumor on diffusion-weighted images because the tumor was too small to be observed on MR images. Of the 48 malignant tumors for which signal intensities could be measured on MR images, the most frequent lesions were squamous cell carcinoma of the uterine cervix (N = 24), adenocarcinoma of the uterine cervix (N = 9), and endometrioid adenocarcinoma of the uterine corpus (N = 5; Table 1). Twelve of the 58 women had typical leiomyomas, for which signal intensities could be measured on diffusionweighted images.

Magnetic resonance imaging protocol
A 3T MR scanner was used (Achieva 3.0T X; Philips Healthcare, Best, The Netherlands) using 32-channel torso/cardiac coils. Unless contraindicated, patients received 20 mg of intramuscular butylscopolamine to prevent artifacts resulting from peristalsis. A parallel imaging technique (SENSE) was used. Axial images of the pelvis were initially obtained using two-dimensional (2D) T2-weighted single-shot turbo spin-echo and three-dimensional T1-weighted gradientecho MR imaging, with and without fat-suppression sequences. The patients then underwent T2-weighted imaging with 2D turbo spin-echo (TSE) sequences. TSE images were obtained at a thickness of 4 mm in both the parasagittal (parallel to the longitudinal axis of the uterus) and axial oblique (orthogonal to the longitudinal axis of the uterus) planes. DWI images were acquired using a single-shot spin-echo echo-planar sequence (b = 0, 700, 1000, 1700, and 2500 s/mm 2 ) in the parasagittal planes (Table 2). We used a maximum b value of 2500 s/mm 2 according to Jensen et al [11]. Dynamic contrast-enhanced images were obtained for all patients, but those images were not used in our study. Table 1. Histopathologic findings of uterine malignant tumors that can be measured on diffusionweighted images.

Histopathologic findings of uterine malignant tumors Number
Squamous cell carcinoma of the uterine cervix 24 Adenocarcinoma of the uterine cervix 9 Poorly differentiated carcinoma of the uterine cervix 3 Small cell carcinoma of the uterine cervix 1 Endometrioid adenocarcinoma of the uterine corpus 5

Image analysis
Using Philips Research Integrated Development Environment (PRIDE) software (Philips Healthcare), we conducted region-of-interest (ROI) measurements of D, K, and the apparent diffusion coefficient (ADC) of malignant tumors of the endometrium and uterine cervix, leiomyoma, healthy myometrium, and healthy endometrium. Two radiologists with different years of experience in abdominal imaging (9 and 5 years) independently drew circular ROIs for each lesion and adjacent non-tumorous regions (healthy myometrium and endometrium) at a b value of 0 s/mm 2 in the region of highest contrast that was devoid of intratumoral necrosis. These ROIs were then automatically copied for other b values (b = 700, 1000, 1700, and 2500 s/mm 2 ). The ROI areas ranged from 12.5 mm 2 to 314.2 mm 2 . D and K were calculated using the following formula: where S and S 0 represent the signal intensities of the images acquired at b and b 0 , respectively (s/mm 2 ). D and K represent diffusivity (mm 2 /s) and kurtosis, respectively. Standard ADC (mm 2 /s) was obtained using a conventional mono-exponential fit with the following equation: where S and S 0 represent the signal intensities acquired at b and b 0 , respectively. DKI parameters (D and K) were calculated using b values ranging from 0 to 2500 s/mm 2 . The ADC was calculated using b values of 0 and 700 s/mm 2 .

Statistical analysis
The intra-class correlation coefficient (ICC) (2,1) was used to assess interobserver agreement between the two radiologists. The mean values for uterine malignant tumor, leiomyoma, healthy myometrium, and healthy endometrium were calculated; the differences in mean values were tested using one-way analysis of variance (ANOVA) with a post-hoc Tukey's honestly significant difference test. An unpaired t-test was used to evaluate the differences in D, K, and ADC between squamous cell carcinoma and adenocarcinoma. The diagnostic accuracy of D and the ADC in differentiating malignant tumors from benign leiomyomas was compared using DeLong's test for two correlated receiver-operating characteristic (ROC) curves. For all statistical analyses other than the ROC analysis, IBM SPSS software (Version 21: IBM, Somers, NY, USA) was used. For ROC analysis, R software (Version 3.2.2: R Foundation for Statistical Computing, Vienna, Austria) was used. A two-tailed P value of <0.05 was considered to indicate a significant difference. Note: Diffusional kurtosis imaging was acquired using a single-shot spin-echo echo-planar sequence. https://doi.org/10.1371/journal.pone.0188434.t002 Diffusional kurtosis imaging in the assessment of uterine tumors: A preliminary study Post-hoc power analysis using the two-sided independent t-test was performed to evaluate the statistical power for detecting significant differences in K between malignant tumors and leiomyomas or between squamous cell carcinomas and adenocarcinomas with the PS software package (version 3.1.2; http://biostat.mc.vanderbilt.edu/wiki/Main/PowerSampleSize).

Results
The DKI and ADC measurements were possible in 48 women with uterine malignant tumors, 12 with leiomyoma, 41 with healthy myometrium, and 19 with healthy endometrium. The ICCs between the two radiologists in evaluating D, K, and the ADCs of the 48 malignant tumors were higher than 0.84, suggesting excellent interobserver agreement ( Table 3). The ICCs for leiomyoma, myometrium, and endometrium ranged from 0.43 to 0.87, suggesting moderate to excellent interobserver agreement (Table 3).  The mean Ds (×10 −3 mm 2 /s) of malignant tumors (1.05 ± 0.41 [standard deviation] and 1.09 ± 0.40 for observers 1 and 2, respectively) were significantly lower than those of leiomyoma (1.40 ± 0.37 and 1.56 ± 0.33, respectively), healthy myometrium (1.72 ± 0.27 and 1.69 ± 0.30, respectively), and healthy endometrium (1.53 ± 0.35 and 1.42 ± 0.37, respectively) for both observers (Fig 4). There were also significant differences between the mean Ds of leiomyoma and healthy myometrium for observer 1 (P = 0.029), and between the mean Ds of healthy myometrium and healthy endometrium for observer 2 (P = 0.039). In comparison, the mean ADCs (×10 −3 mm 2 /s) of malignant tumors (1.00 ± 0.32 and 1.02 ± 0.36 for observers 1 and 2, respectively) were significantly lower than those of leiomyoma (1.34 ± 0.33 and 1.52 ± 0.36, respectively), healthy myometrium (1.55 ± 0.24 and 1.52 ± 0.25, respectively), and healthy endometrium (1.30 ± 0.34 and 1.25 ± 0.28, respectively) for both observers (Fig 5). There were also significant differences between the mean ADCs of healthy myometrium and healthy endometrium for both observers. A comparison of Figs 4 and 5 shows that the mean D and mean ADC had similar distributions for uterine lesions, healthy myometrium, and healthy endometrium. In ROC analysis, there was no significant difference in the area under the ROC Table 3  curve between the mean D and mean ADC in terms of the accuracy of differentiation between malignant tumors and benign leiomyomas for both observers (Fig 6). The mean Ks of malignant tumors (0.88 ± 0.28 and 0.90 ± 0.23 for observers 1 and 2, respectively) were significantly higher than those of healthy myometrium (0.72 ± 0.10 and 0.73 ± 0.10, respectively) and healthy endometrium (0.65 ± 0.13 and 0.60 ± 0.18, respectively) Outliers are also represented (˚and *). The top and bottom of each box represent the 25th and 75th percentiles of the mean ADC, respectively. The horizontal line inside each box represents the median value. The graphs show a significantly lower mean ADC for uterine malignant tumor than for leiomyoma, healthy myometrium, and healthy endometrium. There were no significant differences in the ADC between leiomyoma and myometrium (P = 0.14 and 1.00 for observers 1 and 2, respectively), or leiomyoma and endometrium (P = 0.98 and 0.095, respectively).

D K ADC
https://doi.org/10.1371/journal.pone.0188434.g005 for both observers (Fig 7). No significant differences were observed between the mean Ks of malignant tumors and those of leiomyoma (0.76 ± 0.14 and 0.77 ± 0.16, respectively), the mean Ks of leiomyoma and those of healthy myometrium and healthy endometrium, or the mean Ks of healthy myometrium and those of healthy endometrium, for either observer. The statistical powers for detecting significant differences in K between malignant tumors and leiomyomas were estimated as 0.25 and 0.40 for observers 1 and 2, respectively, suggesting that the sample size was not sufficient.
The Ds, ADCs, and Ks of squamous cell carcinoma of the uterine cervix (N = 24) were compared with those of adenocarcinoma (N = 17) (adenocarcinoma of the uterine cervix, 9; endometrioid adenocarcinoma of the uterine corpus, 5; mixed carcinoma of the uterine corpus, 2; and clear cell adenocarcinoma of the uterine corpus, 1). The mean D of squamous cell carcinomas (0.95 ± 0.18 and 0.98 ± 0.18 for observers 1 and 2, respectively) were significantly lower than those of adenocarcinoma (1.26 ± 0.59 and 1.32 ± 0.54, respectively) for both observers (P = 0.045 and 0.020, respectively). The mean ADCs of squamous cell carcinomas (0.89 ± 0.14 and 0.90 ± 0.16 for observers 1 and 2, respectively) were also significantly lower than those of adenocarcinoma (1.16 ± 0.42 and 1.25 ± 0.49, respectively) for both observers (P = 0.018 and 0.011, respectively). There were no significant differences between the mean Ks of squamous cell carcinoma (0.87 ± 0.22 and 0.89 ± 0.11, respectively) and adenocarcinoma (0.82 ± 0.29 and 0.82 ± 0.26, respectively) for either observer (P = 0.53 and 0.31, respectively). The statistical powers for detecting significant differences in K between squamous cell carcinomas and adenocarcinomas were estimated as 0.08 and 0.13 for observers 1 and 2, respectively.

Discussion
Recently, several scientific studies have focused on DKI to attempt to elucidate the non-Gaussian properties of water diffusion in biologic and pathologic tissues, particularly in cerebrovascular stroke, demyelinating diseases, white matter aging, Parkinson's disease, prostate cancer, breast cancer, and kidney cancer [14,[16][17][18][19][20][21][22][23][24][25][26]. However, the application of DKI for the uterus has not been well investigated. To our knowledge, there is only one recently published report that investigated the utility of various non-mono-exponential analyses of DWI including DKI for uterine cervical tumors [27]. In this study, we successfully performed DKI analysis of the uterus. We found that the mean Ds of cervical and endometrial cancers were significantly lower than those of leiomyoma, healthy myometrium, and healthy endometrium. ROC curve analyses (Fig 6) and a comparison among Figs 4 and 5 demonstrated equality in terms of diagnostic accuracy between D and the ADC.
Regarding the contribution of K, which relates to microstructural complexity, to the evaluation of uterine tumors, the mean K was higher in uterine cancers than in benign tissues such as the myometrium, endometrium, or leiomyoma. These findings suggest that diffusion in malignant uterine lesions shows greater deviation from Gaussian behavior compared with that in benign tissues, which is similar to that in breast and prostate cancers [19,20,[28][29][30]. The differences were statistically significant between uterine cancer and either the myometrium or endometrium, although we were unable to demonstrate a statistically significant difference between cancer and leiomyoma. Because post-hoc power analysis suggested that the sample size was not sufficient, further studies with a larger sample size are definitely needed in terms of the role of K in differentiating between malignant and benign tissues. We also investigated the value of K in discriminating between different pathological types of uterine malignant tumors (i.e., squamous cell carcinoma and adenocarcinoma). There was no significant difference in mean K between the two types of malignant tumors. However, post-hoc power analysis suggested that the sample size was not sufficient; therefore, further studies with a larger sample size are needed for evaluating the role of K in differentiating between these types.
In terms of the DKI measurements, the interobserver agreements were good when the measurements were successful. Therefore, the application of DKI should be clinically feasible. However, it was sometimes impossible to measure the necessary values because of severe artifacts, as shown on Figs 2 and 3. As DWI with a high b value (such as 2500 s/mm 2 ) has a tendency to have reduced signal intensity and stronger artifacts compared with those with lower b-value images, clinical application of the DKI technique may be limited due to the occurrence of stronger artifacts. Improving image quality for high b-value DWI would be crucial for the future use of DKI of the uterus. In addition to the problem of image quality, the longer acquisition time of DKI may be a problem in its routine use. Optimization of scan techniques to reduce acquisition time will also be important from the perspective of practical application [31].
Recently, Winfield et al. reported that DWI data analyses up to b value of 800 s/mm 2 using non-mono-exponential models, such as stretched exponential, kurtosis, statistical, and bi-exponential models, were useful to distinguish between types and grades of uterine cervical tumors [27]. They noted that D was significantly different between tumor grades, and K was significantly different between squamous cell carcinomas and adenocarcinomas. However, as they described, the use of a maximum b value of 800 s/mm 2 was a limitation of their study because the DKI model was originally developed over a much wider range of b values [11,27]. Jensen et al. reported that the precision of DKI estimates rapidly decreased when the maximum b values were substantially reduced below 2000 s/mm 2 in the brain [11]. In our study, we showed the feasibility of DKI analysis up to a higher b value of 2500 s/mm 2 with interobserver reliability for uterine corpus and cervical tumors, leiomyomas, myometrium, and endometrium.
There are several limitations to this study. The first is related to its retrospective nature. The second is related to the limited number of participants, especially for evaluating the role of K. Third, the findings of this study are preliminary; further studies involving several centers and a larger sample size are necessary to better understand the value of DKI in the evaluation of uterine cancers. Fourth, in some cases, we were unable to perform the ROI measurement of the uterine tumor because of difficulties detecting or circumscribing the tumor. Fifth, we did not compare DKI measurements between different kinds of fitting software. Sixth, we did not evaluate the effect of tumor size or location on DKI measurements. Seventh, we did not conduct histogram analysis even though the technique can be useful in characterizing tumors.
In conclusion, the results of this preliminary study on the utility of non-Gaussian DKI in the assessment of uterine tumors suggest that interobserver agreements in evaluating D, K, and ADC are moderate to excellent; D performs equivalent to conventional DWI in differentiating between benign and malignant uterine tumors; and the mean K of malignant uterine lesions is significantly higher than that of non-tumorous myometrium or endometrium.
Supporting information S1 File. Study data set. (XLS)