DTI histogram parameters correlate with the extent of myoinvasion and tumor type in endometrial carcinoma: a preliminary analysis

Abstract

Background

Myoinvasion and tumor-type determines surgical planning in endometrial carcinoma.

Purpose

To evaluate whole tumor diffusion tensor imaging histogram texture parameters in evaluating myoinvasion and tumor type in endometrial carcinoma.

Material and Methods

Twenty-seven patients with endometrial carcinoma underwent diffusion tensor imaging on a 1.5-T MRI system using echo-planar imaging sequence with 0 and 700 s/mm² b values. Whole tumor histogram parameters were obtained from fractional anisotropy, mean diffusivity maps. Mann–Whitney U test and receiver operating characteristic curve analyses were used

Results

The mean fractional anisotropy of tumors with no myoinvasion was significantly higher than tumors which underwent myoinvasion, suggesting higher anisotropy in tumors which did not invade the myometrium. Voxel-wise heterogeneity in distribution of fractional anisotropy and mean diffusivity was seen in the form of higher uniformity and lower entropy of tumors with superficial <50% myoinvasion versus >50% myoinvasion. Uniformity, entropy, and energy of voxel-wise fractional anisotropy distribution gave an area under the curve of 0.827, 0.821, and 0.796, respectively, in predicting the presence of deep myometrial invasion while energy, entropy, and uniformity of mean diffusivity distribution in tumor gave an area under the curve of 0.84, 0.815, and 0.809 respectively. Tumor type was predicted with an area under the curve of 0.747, 0.759, and 0.765 for the uniformity, energy, and entropy of voxel-wise fractional anisotropy distribution. A logistic regression combining all the important histogram parameters obtained 94% and 88% sensitivity and 88% and 80% specificity in predicting deep myoinvasion and tumor type, respectively.

Conclusion

Diffusion tensor histogram analysis can better characterize endometrial carcinomas and can be used as a quantitative marker of tumor behavior.

Keywords

Diffusion tensor imaging endometrial carcinoma staging histogram analysis texture analysis

Introduction

Endometrial carcinoma, the most common gynecological malignancy in the western hemisphere is frequently treated surgically with good patient outcomes. Depth of myoinvasion correlates with five-year patient survival rates and the prevalence of lymph nodal metastases (1 –4). Tumor grade also affects survival and type 2 endometrial tumors have a higher incidence of myometrial, cervical stromal, and lymphovascular stromal invasion compared to type 1 tumors. Pelvic lymph nodal dissection is commonly limited to either type 2 endometrial tumors or to type 1 tumor with deep (>50%) myoinvasion. Preoperative contrast-enhanced dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is the gold standard for the delineation of the depth of myoinvasion, which determines further surgical planning (5). Contrast administration, however, is not without its disadvantages and diffusion-weighted imaging (DWI) is an alternative imaging modality used in preoperative imaging. Qualitative evaluation of DWI can delineate the myometrial tumor interface when: (i) contrast administration is contraindicated due to its associated risks (6, 7); (ii) when tumoral and myometrial enhancements are similar; and (iii) when the myometrium is effaced by atrophy or due to large fibroids. The quantitative evaluation of ADC maps in endometrial carcinomas, including filtration histogram texture analysis has been shown to predict the extent of myoinvasion and tumor type of endometrial tumors (8, 9).

DWI assumes that the restriction of diffusion is similar in all directions. This assumption is, however, an oversimplification in both neurological and general visceral imaging. With a fractional anisotropy (FA) of 0.297 in the junctional zone, 0.257 in the outer myometrium, and 0.186 in normal endometrium, the diffusion is more anisotropic in the myometrium as compared to the endometrium (10). Zhang et al. demonstrated that the FA values of diseased superficial myometrium with tumor invasion were different from those of the normal superficial myometrium (11). The FA values are also significantly different in various grades of endometrial carcinoma (12). Since FA is a measure of anisotropy, we hypothesize that differences in the degree of anisotropy of the tumor can also reflect tumor myoinvasion or tumor type.

Given this background, we evaluate whether whole tumor diffusion tensor imaging (DTI) histogram texture parameters varied in groups having a superficial and deep myometrial invasion and in type 1 and 2 endometrial tumors.

Material and Methods

Patients

Approval was obtained from the Institutional Review Board and written consent was obtained from every enrolled patient. The patients were selected sequentially from those referred by our institutional gynaecological outpatient department (OPD) as curettage proved cases of endometrial malignancy for preoperative staging pelvic MRI. Patients having a contraindication for contrast administration or those not being taken up for surgery (because of advanced disease) were excluded from the study. None of the patients included in the study had received neoadjuvant chemotherapy. The menstrual phase of the patients was not recorded before imaging. A total of 31 patients were referred for inclusion into the study, of whom one patient refused imaging because of claustrophobia. Another three patients had undergone DCE-MRI before referral and surgical planning was done without repeat imaging. Finally, 27 consecutive patients were prospectively enrolled between July 2015 to December 2016 and informed written consent was obtained. The general demographics of the patients included in the current study are given in Table 1.

Table 1.

Demographics of the cases included in the study.

	Cases (n)	Total (%)
Cases with no superficial myoinvasion	5	18.52
Cases with superficial myoinvasion	13	48.15
Cases with deep myoinvasion	9	33.33
Histological subtype
Endometroid adenocarcinoma	19	70.37
Grade 1	7	25.93
Grade 2	10	37.04
Grade 3	2	7.41
Serous carcinoma	3	11.11
Others	5	18.52
Type of tumor
Type 1	17	62.96
Type 2	10	37.04
Cases with endocervical extension	4
Cases with cervical stromal invasion	2

MR system and scanning methods

All patients were scanned using a Siemens Aera 1.5-T scanner with syngo MR D13 software version with a 16-phased body array coil. The bladder was partially distended to ensure minimal patient discomfort or motion. A HASTE localizer, T2-weighted (T2W) images in three orthogonal planes perpendicular to the long axis of the uterus, DWI in the sagittal plane, pre-contrast T1-weighted (T1W) fat-saturated images in the axial plane, DCE images, and post-contrast T1W fat-saturated images in the axial and sagittal planes were acquired.

Single-shot echo planar DTI was acquired in the sagittal plane with a resolution matrix of 128 × 104; 30 slices were obtained, with a field of view (FOV) read of 250 mm and FOV phase of 81.3%; a 5-mm slice thickness with slice gap of 20% was used; the TR and TE were 4900 ms and 73 ms, respectively. The fat saturation used was weak with subtraction being used for flow compensation and water suppression. B0 images with number of excitations (NEX) 8 and multidirectional DW images in 12 different diffusion directions with NEX 5 were acquired using the bipolar diffusion scheme. GRAPPA, with an acceleration factor of 2, was used for parallel imaging to reduce acquisition time.

Post-processing of the diffusion tensor images with quantitative assessment

The images were exported in the DICOM format and were opened on Horos version 2.0.1 in a Mac Book Pro (Processor 2.7 GHz, Intel Core i5, Retina, 13-inch). The trace diffusion tensor images were opened using a four-dimensional (4D) viewer and an open source plugin DTIMap was used to derive FA (dimensionless) (×10⁻⁴) and mean diffusivity (MD/ADC) (× 10⁻⁶ mm²/s) parametric maps. The maps were a fit to $S_{i} = e^{(- b_{i} {GDG}^{'})}$ where $S_{i}$ is the images at B values $b_{i}$ , G is the diffusion gradient direction, and D is the 3 × 3 diffusion tensor. The diffusion tensor D is then decomposed to produce eigenvector and eigenvalues. The fitting and decomposition are performed using the linear algebra package (LAPACK) routines. In short (13, 14), anisotropic diffusion in each voxel of tissue can be characterized in the form of a diffusion ellipsoid, whose main axis is along the principal direction of diffusion restriction in the voxel. $λ_{1}$ is the value of the unit vector $ε_{1}$ in the principal diffusion direction representing the axial or parallel diffusivity, while λ_2, λ₃ are orthogonal to ε₁ such that λ₂ > λ₃ this represents perpendicular diffusivity. The other two scalar parameters used are as follows:

\frac{λ_{1} + λ_{2} + λ_{3}}{3} = Mean diffusivity (MD)

\begin{array}{l} \frac{1 \sqrt[2]{{(λ_{1} - λ_{2})}^{2} + {(λ_{1} - λ_{3})}^{2} + {(λ_{2} - λ_{3})}^{2}}}{\sqrt[2]{2} \sqrt[2]{{(λ_{1})}^{2} + {(λ_{2})}^{2} + {(λ_{3})}^{2}}} \\ = Fractional Anisotropy (FA) \end{array}

Data collection: histogram analysis

FA, MD maps, and the b0 maps for each case were exported and saved in the DICOM format. 3D Slicer 4.11.0 (https://download.slicer.org/) was used for defining the region of interest (ROI) and for calculation of the texture parameters. Two radiologists with five and fifteen years of experience, working in consensus, drew a freehand ROI on the trace b0 images of endometrial tumor excluding areas of hemorrhage and necrosis after selecting the slice with the maximum tumor bulk and adjoining 3–4 slices depending on the size of the tumor. DCE images and T2 images were consulted during the drawing of ROIs to ensure necrosis and hemorrhage were adequately accounted for. The b0 images were used because they provided adequate anatomical information (Fig. 1). No evaluation was done regarding the depth of myoinvasion during the drawing of the ROIs and the peripheral part of a tumor at the tumor–myometrial junction was not included to avoid erroneous inclusion of the myometrium. The ROIs were then saved as a label map which was used to obtain the following first order histogram statistics (seven features) using the pyRadiomics plugin (https://pyradiomics.readthedocs.io/en/latest/index.html): skewness; uniformity; energy; entropy; variance; kurtosis; and mean. The first order histogram-based texture features are intuitive and are easy to interpret. The parameters are concerned with the distribution of the signal intensity in the individual voxels without involving itself with the spatial distribution of the voxels. Uniformity refers to the consistency of the signals in all the voxels of the ROI, while entropy refers to the randomness of the distribution of the signal in the voxels. The asymmetry and the flatness of the histogram are referred to by the skewness and kurtosis, respectively. First order-based energy refers to the magnitude of voxel values in the image: it is volume-dependent and a higher value demonstrates a higher sum of squares of the voxel intensities (15).

Fig. 1.

For the generation of the whole tumor histogram parameters, the b0 map was opened in 3D slicer and a label map was drawn on the slice and the adjoining 3–4 slices with the most substantial bulk of the tumor. Areas of hemorrhage and necrosis were excluded (dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and T2 images were used as reference) and the peripheral myometrial–tumor interface was avoided to prevent erroneous inclusion of the myometrium. Subsequently, this label map was used to calculate the histogram parameters from the fractional anisotropy (FA) and mean diffusivity (MD/ADC) maps.

Fig. 2.

Box-and-whisker plot showing the distribution of the various whole tumor histogram parameters which were statistically different in groups of cases having >50% myoinvasion and in the group with no or <50% myoinvasion.

Visual interpretation by radiologists

Visual interpretation by radiologists was done six months prior, in order to avoid any observer bias associated with drawing the ROIs. The cases were anonymized and interpreted separately by both the radiologist who observed T2W, DWI, and DCE-MRI and interpreted the depth of myometrial and cervical invasion. The DCE images were used for clinical decision-making and while drawing of ROIs to exclude necrosis; however, the final interpretation of the depth of myoinvasion for this study was obtained using histopathology.

Surgery and histopathology

All the patients included in the study underwent total abdominal hysterectomy with bilateral salipingo-oophorectomy. The tumour histology, grade, lymphovascular stromal invasion, presence, depth of myometrial invasion, extension to the serosa and the cervix, cervical stromal invasion, the involvement of the tube, ovaries, and presence of malignant cells in peritoneal washings were recorded.

Statistical analysis

The demographics, imaging findings, and the various histogram parameters obtained from FA and MD maps were tabulated and recorded in IBM SPSS Statistics v.23.0. The continuous texture variables were tested for normality using the Shapiro–Wilk test and were found to significantly deviate from a normal distribution. Mann–Whitney U test, a non-parametric test, was used to test whether texture parameters across the different groups were significantly different. A P value < 0.05 was considered statistically significant. Receiver operating characteristic (ROC) curve analysis was used to obtain the area under the curve (AUC) for texture parameters found to be significantly different between the various groups; optimal cut-offs were obtained using bootstrapped Youden index. A logistic regression classifier was built using all the statistically significant texture parameters and a 10-fold cross-validation was done to evaluate the generalizability of the developed model. Box-and-whisker plots and ROC curves were made using Medcalc software v.14.8.1.

Results

There was a statistically significant difference in mean FA between the group having no myoinvasion (n = 5; 2418.441 ± 523.388) versus the group with myoinvasion (n = 22; 1817.858 ± 629.061) (P < 0.05 using the Mann–Whitney U test). A ROC curve analysis showed an AUC of 0.791(95% confidence interval [CI] = 0.592–0.922) with an optimal cut-off of ≤2156.239 (95% CI > ≤1585.149–2561.420) yielding 81.82% sensitivity (95% CI = 59.7–94.8) and 80% specificity (95% CI = 28.4–99.5). The mean MD (943.252 ± 316.335) was lower in cases with myoinvasion versus cases without myoinvasion (1169.599 ± 194.672); however, this was not statistically significant. The mean FA and mean MD were not significantly different between groups of type 1 and type 2 endometrial tumors.

Tumors having myoinvasion limited to the superficial 50% had a statistically significant lower energy and entropy and a higher uniformity of FA distribution versus tumors with >50% myoinvasion (Fig. 2 and Table 2). Uniformity of FA provided an AUC of 0.827 (95% CI = 0.633–0.944) with a cut-off of ≤0.0244 (95% CI ≤0.0214–0.0244) having a sensitivity of 100% (95% CI = 66.4–100.0) and specificity of 66.67% (95% CI = 41.0–86.7) in predicting deep myoinvasion. The energy of FA had an AUC of 0.796 (95% CI = 0.598–0.926) with a cut-off of >489,461,973 (95% CI = 119, 233,392–623,941,252) having 88.89% sensitivity (95% CI = 51.8–99.7) and 72.22% specificity (95% CI = 46.5–90.3) in predicting deep myometrial invasion (Fig. 4 and Table 3). The group having superficial myoinvasion has lesser skewness, energy, entropy, variance, and kurtosis and greater uniformity of the distribution of MD than cases with deep myoinvasion (Table 2). The entropy of MD provided an AUC of 0.815 (95% CI = 0.619–0.937) while skewness of MD had an AUC of 0.759 (95% CI = 0.557–0.902) in predicting deep myometrial invasion (Table 3).

Table 2.

Histogram texture parameters obtained from the fractional anisotropy (FA) and mean diffusivity (MD) maps compared between groups with <50% and >50% myoinvasion and groups with type 1 and 2 endometrial tumors.

	Deep myoinvasion (mean (SD))			Tumor type (mean (SD))
	Absent	Present	P value*	Type 1	Type 2	P value^*
Mean diffusivity
Skewness	0.688(0.756)	1.838(1.227)	0.031	0.965(0.988)	1.253(1.228)	0.651
Uniformity	0.131(0.21)	0.049(0.011)	0.01	0.072(0.023)	0.158(0.286)	0.098
Energy	167,617,880.5 (309,421,422.806)	333,746,215.11 (297,496,852.136)	0.005	115,307,454.18(149,143,099.01)	406,061,106.4(423,354,500.785)	0.098
Entropy	3.859(1.225)	4.721(0.389)	0.009	4.131(0.525)	4.174(1.724)	0.079
Variance	38,607.681 (41,989.348)	90,651.619 (89,139.22)	0.016	34,386.674(26,201.229)	92,622.937(92,767.872)	0.108
Kurtosis	4.057(1.723)	9.202(6.518)	0.012	5.302(3.826)	6.571(5.82)	0.547
Mean	964.422(360.696)	1026.659(168.362)	0.918	1036.104(182.973)	898.576(448.08)	0.547
Fractional anisotropy
Skewness	0.514(0.463)	0.848(0.481)	0.15	0.617(0.534)	0.641(0.422)	0.451
Uniformity	0.082(0.215)	0.016(0.005)	0.006	0.03(0.017)	0.111(0.292)	0.035
Energy	528,576,933.22 (580,081,304.447)	1,371,620,035.89(1,354,034,055.992)	0.014	437,149,795.71(370,414,831.612)	1,442,741,859.4(1,337,068,648.769)	0.027
Entropy	5.106(1.464)	6.213(0.438)	0.007	5.39(0.732)	5.621(2.011)	0.024
Variance	500,742.58 (266,363.465)	682,291.896(435,632.078)	0.504	619,106.679(339,602.762)	462,917.995(320,282.286)	0.269
Kurtosis	3.565(1.13)	4.514(2.217)	0.537	3.846(1.751)	3.941(1.387)	0.451
Mean	1858.848(650.944)	2069.537(652.458)	0.681	2057.375(573.996)	1710.972(733.959)	0.366

*Based on Mann–Whitney U non-parametric test.

Table 3.

ROC curve analysis of the various statistically significant histogram texture parameters obtained from the FA and MD maps in differentiating cases of superficial and deep myometrial invasion.

Variable	Area under the ROC curve (95% CI)	Associated criterion*	Sensitivity (%)	Specificity (%)
FA
Entropy	0.821 (0.626–0.941)	> 5.525 (>5.4771–>6.5435)	100 (66.4–100.0)	66.67 (41.0–86.7)
Uniformity	0.827 (0.633–0.944)	≤0.0244 (≤0.0214–≤0.0244)	100 (66.4–100.0)	66.67 (41.0–86.7)
Energy	0.796 (0.598–0.926)	>489,461,973 (>119,233,392–>623,941,252)	88.89 (51.8–99.7)	72.22 (46.5–90.3)
MD
Energy	0.84 (0.648–0.951)	>75,479,276 (>59,079,944–>116,789,049)	100 (66.4–100.0)	72.22 (46.5–90.3)
Entropy	0.815 (0.619–0.937)	>4.076 (>3.9847–>4.2717)	100 (66.4–100.0)	66.67 (41.0–86.7)
Kurtosis	0.802 (0.605–0.929)	>3.5417 (>3.5257–>7.3658)	100 (66.4–100.0)	55.56 (30.8–78.5)
Skewness	0.759 (0.557–0.902)	>1.5999 (>0.1802–>1.9365)	55.56 (21.2–86.3)	94.44 (72.7–99.9)
Uniformity	0.809 (0.612–0.933)	≤0.0595 (≤0.0497–≤0.067)	88.89 (51.8–99.7)	77.78 (52.4–93.6)
Variance	0.79 (0.591–0.922)	>22,631.14 (>20,109.4242–>28,548.2183)	100 (66.4–100.0)	61.11 (35.7–82.7)

The optimal cut-off was derived to maximize the average of sensitivity and specificity.

*Cut-off based on bootstrapped Youden index.

ROC, receiver operating characteristic; FA, fractional anisotropy; MD, mean diffusivity; CI, confidence interval.

The 10-fold cross-validated logistic regression model combining the significant texture parameters deep myoinvasion with 94.44% sensitivity (95% CI = 72.71–99.86), 88.89% specificity (95% CI = 51.75–99.72), and an AUC of 0.92 (95% CI = 0.74–0.99) (Table 5).

Table 4.

ROC curve analysis of the various statistically significant histogram texture parameters obtained from the FA and MD maps in differentiating cases of with type 1 and type 2 endometrial tumors.

FA
Variable	AUC (95% CI)	Associated criterion*	Sensitivity (%) (95% CI)	Specificity (%) (95% CI)
Uniformity	0.747 (0.544–0.893)	≤0.0176 (≤0.016–≤0.0769)	80 (44.4–97.5)	88.24 (63.6–98.5)
Energy	0.759 (0.557–0.901)	>665045949 (>623941252–>1480668687)	80 (44.4–97.5)	88.24 (63.6–98.5)
Entropy	0.765 (0.563–0.905)	>6.0331 (>5.9277–>6.5311)	80 (44.4–97.5)	88.24 (63.6–98.5)

The optimal cut-off was derived to maximize the average of sensitivity and specificity.

*Cut-off based on bootstrapped Youden index.

ROC, receiver operating characteristic; FA, fractional anisotropy; MD, mean diffusivity; AUC, area under the curve; CI, confidence interval.

Table 5.

Accuracy metrics of the logistic regression model built to combine all the texture features which were found to be statistically significant in predicting the presence of deep myometrial invasion and type of tumor.

	Deep myometrial invasion (95% CI)	Type of tumor (95% CI)
Sensitivity (%)	94.44 (72.71–99.86)	88.24 (63.56–98.54)
Specificity (%)	88.89 (51.75–99.72)	80.00 (44.39–97.48)
AUC	0.92 (0.74–0.99)	0.84 (0.65–0.95)
Positive likelihood ratio	8.5 (1.33–54.13)	4.41 (1.26–15.42)
Negative likelihood ratio	0.06 (0.01–0.43)	0.15 (0.04–0.56)
Disease prevalence (%)	66.67 (46.04–83.48)	62.96 (42.37–80.60)
PPV	94.44 (72.71–99.86)	88.24 (63.56–98.54)
NPV	88.89 (51.75–99.72)	80.00 (44.39–97.48)

The model was built using 10-fold cross-validation whereby the data were randomly split into 10 folds: one fold was used for validation, one for testing, and the rest were used to train the model. This was repeated 10 times and the averaged accuracy metrics over the validation dataset are provided.

CI, confidence interval; AUC, area under the curve; PPV, positive predictive value; NPV, negative predictive value.

Uniformity, energy, and entropy of FA were all significantly lower in type 1 tumors compared to type 2 tumors (Fig. 3 and Table 2). Uniformity of FA for type 1 tumors was 0.03 ± 0.017 compared to 0.111 ± 0.292 for type 2 tumors with an AUC of 0.747 (95% CI = 0.544–0.893). Entropy provided an AUC of 0.765 (95% CI = 0.563–0.905) (Fig. 5 and Table 4).

The 10-fold cross-validated logistic regression model to predict tumor type had a sensitivity of 88.24% (95% CI = 63.56–98.54) and a specificity of 80.00% (95% CI = 44.39–97.48), with an AUC of 0.84 (95% CI = 0.65–0.95) (Table 5). Thus, we were able to demonstrate that anisotropic diffusion can better characterize endometrial carcinomas.

Discussion

FA is a proxy of the asymmetry of diffusion in tissues. DTI demonstrates the degree of disorganization of tissue architecture in malignancies and has been used to evaluate the depth of invasion in esophageal and gastric carcinomas (16, 17). When diffusion is similar in all directions, FA equals zero. FA equals 1 when diffusion restriction is in one direction. We saw that the mean FA (anisotropy) of tumors without myoinvasion was higher (2418.441 ± 523.388) compared to tumors with myoinvasion (1817.858 ± 629.061) in our study. This may be explained by the fact that endometrial tumors which do not invade the myometrium are less aggressive and will have a better preservation of the glandular architecture (Fig. 6). This causes a more significant diffusion restriction in one particular direction and higher anisotropy. The loss of the orderly architecture in aggressive tumors with invasion would result in similar diffusion restrictions in all directions and hence higher isotropy.

Fig. 3.

Box-and-whisker plot showing the distribution of the various whole tumor histogram fractional anisotropy (FA) parameters which were statistically different in type 1 and type 2 endometrial tumors.

Fig. 4.

Receiver operating characteristic (ROC) curves of the histogram parameters found to be significant in predicting the depth of myoinvasion.

Fig. 5.

Receiver operating characteristic (ROC) curves of the histogram parameters found to be significant in predicting the type of tumor.

Fig. 6.

(a) Well-differentiated endometrioid carcinoma (type I carcinoma) showing preserved glandular architecture lined by tall columnar well-differentiated tumor cells. (b) Serous uterine carcinoma (type II carcinoma). Irregular glands with poorly maintained architecture are seen lined by pleomorphic tumor cells with high-grade nuclei infiltrating the myometrium. The inset shows tumors at high magnification with papillae lined by pleomorphic tumor cells. Hematoxylin and eosin stain, ×40, inset ×200. We hypothesize that in better-differentiated tumors, the preserved architecture results in the predominance of diffusion restriction in one particular direction versus higher grade tumors where tissue disorganization may result in equivalence of diffusion in all directions.

Texture characterizes voxel-to-voxel heterogeneity of tumors and provides quantitative descriptors of heterogeneity. Texture analysis demonstrated the heterogeneity in the voxel-wise distribution of apparent diffusion coefficients (ADC) in endometrial carcinomas (8, 9). Ytre-Hauge et al. demonstrated that an entropy of >4.49 for distribution of ADCs in ROIs drawn on the tumor (with Laplacian of Gaussian filter) correlated with the presence of >50% myometrial invasion. This provided 70% sensitivity, 84% specificity, and 78% overall accuracy. High-grade tumors, which commonly undergo deeper myoinvasion, are heterogeneous with microscopic areas of necrosis, hemorrhage, and region-to-region variability of tissue architecture. This would translate into a greater voxel-to-voxel variability of ADCs and in our study variability of FA and MD values. This variability in tumor diffusion is picked up by texture analysis of the FA and MD maps in the form of greater randomness (entropy) and lesser uniformity of the voxel-wise distribution of FA and MD values of tumors which undergo deep myoinvasion (Fig. 2). This heterogeneity is also quantified by lower skewness, entropy, variance, and kurtosis of the distribution of MD values in the tumor limited to the superficial 50% of the myometrium versus tumors with >50% myoinvasion.

Interestingly, the energy FA and energy MD were higher in the cases having >50% myoinvasion. In our study, whole tumor ROIs were used; since energy is dependent on the size of the ROI, energy may simply reflect the fact that the tumors that invaded >50% of the myometrium were generally bulkier and larger in size. This translated into a larger ROI and hence larger energy values.

In our study, FA of grade 1 endometrioid adenocarcinoma was 2187.787 ± 509.768 compared to 1972.415 ± 569.816 for grade 2 tumors and 1605.696 ± 29.057 for grade 3 tumors. The findings are consistent with the findings of Yamada et al. (12) who demonstrated that FA of grade I tumors is higher compared to grade 2 and 3 tumors. Ytre-Hauge et al. had demonstrated that high-risk type 2 endometrial tumors had a higher entropy of the voxel-wise distribution of ADC at 4.6 (95% CI = 4.3–4.9) compared to type 1 tumors at 4.0 (95% CI = 3.9–4.3). We also found similar results for the entropy of FA distribution with type 2 tumors having a higher entropy (5.621 ± 2.011) compared to type 1 tumors (5.39 ± 0.732) (Fig. 3).

We are not the first to use whole tumor histogram texture parameters obtained from FA and MD maps. FA texture features have been used in the classification of gliomas to identify IDH1 mutation and 1p/19q codeletion status (15) and to classify various brain tumors and white matter diseases (9, 18–20). While studies have utilized DTI for staging endometrial carcinomas, texture analysis of DTI remains a novel approach in the endometrium. Yamada et al. demonstrated that the FA of the junctional zone is significantly different from the deep myometrium and endometrial tumors. Based on the above, they suggested a visual interpretation of the parametric DTI maps to predict the depth of myoinvasion. However, DTI parametric maps are highly pixelated and this would interfere with the visual interpretation, especially when the myometrium is thinned out in postmenopausal women or when the myometrium is distorted by myomas or adenomyosis. Zhang et al. (11) demonstrated that the FA and ADC of the superficial myometrium having tumor invasion is significantly lower than the normal uninvolved superficial myometrium. However, since the above model involves drawing ROIs on the subjectively perceived normal and abnormal myometrium, its implementation in routine clinical practice will be subject to greater inter-observer variability. We assume the whole tumor ROI-based texture approach will be easier and will have a greater reproducibility adding to radiologists’ confidence. Our study had two additional advantages over the texture-based model described by Ytre-Hauge et al. First, we use an open-source graphical user interface based easily accessible software to derive the texture parameters, which means that our results can be tested for reproducibility, generalizability, and implementation without adding significant cost burden to clinical care. Second, Ytre-Hauge et al. evaluated texture parameters obtained from multiple image sequences which were not limited to ADC maps only but also T2 images and post-contrast sequences. We, on the other hand, limit ourselves to FA and MD maps obtained from a single 3-min imaging sequence. We hope by further expanding on our preliminary findings, we can develop FA texture parameter-based models which would determine the presence of tumor invasion, tumor type, and triage the requirement for administration of during MR imaging.

Our study has two significant limitations. First, we have a relatively small sample size. However, we consider our findings preliminary and the results need to be further validated in a more extensive study. Another problem, though not intrinsic to our research, is the standardization of DTI. Our study utilized 12 diffusion-sensitizing directions, which is different from the other studies evaluating the uterus. Different post-processing algorithms used for parametric map generation may result in a difference in the generated DTI parameters. These may interfere with the generalizability of the study; however, production of institution-wise datasets, imaging protocols, and cut-offs may be a workaround.

In conclusion, whole tumor histogram parameters obtained from FA and MD maps can predict the type of tumor and the extent of myoinvasion in endometrial carcinoma. We believe these parameters will serve as imaging markers of tumor invasion and type. Development of classification machine-learning algorithms utilizing texture parameters obtained from FA and MD maps may allow highly accurate prediction of tumor type and extent of myometrial invasion which will help decide further administration of contrast. More extensive studies with standardization of DTI acquisition protocols are needed for future evaluation and incorporation of DTI in routine uterine carcinoma MRI protocols.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

References

Todo

Kato

Kaneuchi

, et al. Survival effect of para-aortic lymphadenectomy in endometrial cancer (SEPAL study): a retrospective cohort analysis. Lancet 2010; 375:1165–1172.

Benedetti Panici

Basile

Maneschi

, et al. Systematic pelvic lymphadenectomy vs. no lymphadenectomy in early-stage endometrial carcinoma: randomized clinical trial. J Natl Cancer Inst 2008; 100:1707–1716.

Kitchener

Swart

Qian

, et al. Efficacy of systematic pelvic lymphadenectomy in endometrial cancer (MRC ASTEC trial): a randomised study. Lancet 2009; 373:125–136.

Huang

Gebb

Einstein

, et al. Accuracy of preoperative endometrial sampling for the detection of high-grade endometrial tumors. Am J Obstet Gynecol 2007; 196:243.e1–5.

Brandao

Silva

AO.

Diseases of the female pelvis: advances in imaging evaluation.

Magn Reson Imaging Clin N Am 2013; 21:447–469.

Beddy

Moyle

Kataoka

, et al. Evaluation of depth of myometrial invasion and overall staging in endometrial cancer: comparison of diffusion-weighted and dynamic contrast-enhanced MR imaging. Radiology 2012; 262:530–537.

Andreano

Rechichi

Rebora

, et al. MR diffusion imaging for preoperative staging of myometrial invasion in patients with endometrial cancer: a systematic review and meta-analysis. Eur Radiol 2014; 24:1327–1338.

Ueno

Forghani

Endometrial carcinoma: MR imaging-based texture model for preoperative risk stratification-a preliminary analysis.

Radiology 2017; 284:748–757.

Ytre-Hauge

Dybvik

Lundervold

, et al. Preoperative tumor texture analysis on MRI predicts high-risk disease and reduced survival in endometrial cancer. J Magn Reson Imaging 2018; 48:1637–1647.

10.

Fujimoto

Kido

Okada

, et al. Diffusion tensor imaging (DTI) of the normal human uterus in vivo at 3 Tesla: comparison of DTI parameters in the different uterine layers. J Magn Reson Imaging 2013; 38:1494–1500.

11.

Zhang

Liu

Zhang

, et al. Use of diffusion tensor imaging in assessing superficial myometrial invasion by endometrial carcinoma: a preliminary study. Acta Radiol 2015; 56:1273–1280.

12.

Yamada

Wakana

Kobayashi

, et al. Endometrial carcinoma: Evaluation using diffusion-tensor imaging and its correlation with histopathologic findings. J Magn Reson Imaging 2019; 50:250–260.

13.

Alexander

Lee

Lazar

, et al. Diffusion tensor imaging of the brain. Neurotherapeutics 2007; 4:316–329.

14.

Soares

Marques

Alves

, et al. A hitchhiker’s guide to diffusion tensor imaging. Front Neurosci 2013; 7:31.

15.

Park

Han

Ahn

, et al. Whole-tumor histogram and texture analyses of DTI for Evaluation of IDH1-mutation and 1p/19q-codeletion status in World Health Organization grade II gliomas. AJNR Am J Neuroradiol 2018; 39:693.

16.

Yamada

Hikishima

Miyasaka

, et al. Esophageal carcinoma: ex vivo evaluation with diffusion-tensor MR imaging and tractography at 7 T. Radiology 2014; 272:164–173.

17.

Yamada

Hikishima

Miyasaka

, et al. Diffusion-tensor MRI and tractography of the esophageal wall ex vivo. J Magn Reson Imaging 2014; 40:567–576.

18.

Nave

Foresti

Pratesi

, et al. Whole-brain histogram and voxel-based analyses of diffusion tensor imaging in patients with leukoaraiosis: correlation with motor and cognitive impairment. AJNR Am J Neuroradiol 2007; 28:1313.

19.

Steffen-Smith

Sarlls

Pierpaoli

, et al. Diffusion tensor histogram analysis of pediatric diffuse intrinsic pontine glioma. Biomed Res Int 2014; 2014:647356.

20.

Wagner

Narayan

Bosemani

, et al. Histogram analysis of diffusion tensor imaging parameters in pediatric cerebellar tumors. J Neuroimaging 2016; 26:360–365.