**4. Texture analysis**

#### **4.1 Technical considerations**

Sometimes the classic imaging features of different types of adnexal lesions may be subtle or overlap, resulting in experts giving the wrong interpretation [25]. For these reasons, in most cases, a definitive diagnosis of an ovarian mass can be made only based on pathological analysis, which raises the patients' risks and healthcare costs.

It is theorized that the several micro and macroscopic histological characteristics of ovarian masses can also be reflected into the background of medical images, but their influence is too subtle to be assessed by the common visual evaluation. Textures represent the intrinsic and intuitive properties of surfaces such as roughness, granulation, and regularity. Texture analysis (TA) is an image processing method based on the extraction and analysis of image-specific parameters that reflect the pixels' distribution patterns and intensity variations [26]. Through these processes, TA provides an objective description of image content by attributing values to several

classes of texture parameters. The basic principle of image-based TA is that a pathological process that alters a tissue produces a modified signal, which will, in turn, give textural features different values from those of the normal structure [27].

The standard TA workflow consists of five steps: image segmentation, feature extraction, feature reduction, feature selection, and class prediction. Image segmentation can be performed automatically, semi-automatically, or manually. It consists of incorporating a given structure into a region of interest (ROI). Most often, researchers choose a semi-automatic technique, where a seed is defined near the center of the target lesion and the software automatically delineates the rest of the lesion based on gradient and geometry coordinates. The region or volume of interest could be delineated as a two or three-dimensional structure, the latter being able to provide more information at the cost of higher definition times [19]. However, it was proven that the latter reduces operator variability associated with multislice/volumetric analysis [20]. Therefore, it remains debatable which form will be best suited for clinical implementation. The TA parameters' extraction is performed automatically in almost all available software. The user can, however, adjust several settings regarding the parameters' computation methods (such as the number of bits/pixel and inter-pixel distances). Most software solutions allow the extraction of a large number of texture features (parameters), which can be more difficult to process by researchers that are not familiar with statistical analysis. Therefore, in order to identify the parameters that are most suited to discriminate between groups, several reduction techniques can be available (such as Fisher, Mutual Information, and the probability of classification error and average correlation coefficients) [21]. The number of parameters can be furtherly reduced by univariate analysis (typically the Mann-Whitney U test), or this analysis could be used as the only reduction and selection technique [28]. Class prediction (or the ability of previously-selected texture parameters to distinguish between vectors belonging to different pre-defined groups) can be performed statistically through the receiver operating characteristic analysis or by the use of classifiers (such as k-nearest neighbors algorithm or artificial neural networks) [29]. One of the most often used software for TA-related medical imaging research remains MaZda, which provides build-in functions for feature selection and class prediction [30]. A typical workflow for feature extraction in MaZda is displayed in **Figure 4**.

#### **Figure 4.**

*Workflow within the MaZda software for extracting texture parameters from a T2-weighted MRI image of a patient with an ovarian hemorrhagic cyst. (A) The region of interest (ROI) seed placed by the researcher (red ellipse) and (B) the ROI that was automatically defined by the software. Before feature extraction, parameter settings can be adjusted (C). Histogram representation based on the parameters extracted from the lesion (D).*

*Quantitative Imaging Parameters in the Diagnosis of Endometriomas DOI: http://dx.doi.org/10.5772/intechopen.101561*


**Table 3.**

*A brief description of the MRI-based texture parameters that showed statistically significant results when comparing endometriomas to HCs, based on the study conducted by Lupean et al. [25].*

#### **4.2 MRI-based TA for the diagnosis of endometriomas**

One study [25] analyzed the texture parameters' ability to distinguish endometriomas from HCs. This study [25] extracted texture features from the internal content of these lesions as seen on T2W. Fourteen parameters showed statistically significant results when comparing the two entities: two variations of the wavelet energy, seven variations of entropy, three of the angular second moment, one of the sum entropy, and the histogram's variance (**Table 3**). Their combined ability was able to differentiate the two entities with a sensitivity of 100% [95% confidence interval (95% CI), 85.8–100%] and a specificity of 100% (95% CI, 71.5–100%).

HCs displayed lower values of wavelet energy, most likely because their content is more homogeneous, resulting in lower signal variation rates [25]. Also, these lesions expressed lower values of the entropy and sum entropy parameters, probably because they do not contain such diverse cell populations and heterogeneous biochemical components compared to endometriomas [36]. Endometriomas on the other hand showed lower values of the angular second moment and higher value of the variance parameter, indicating a lesser uniform content for these lesions (**Figure 5**). These parameters were considered reflections of the heterogeneous content of endometriomas, which otherwise could not be assessed by the usual examinations of the MRI images [25].

**Figure 5.** *(A) T2-weighted image of a patient with histologically-proven endometriomas. The texture maps show the distribution of the variance (B) and wavelet energy (C) parameters.*

## **4.3 Ultrasound-based TA for the diagnosis of endometriomas**

The ultrasound (US) appearance of endometriomas mainly depends on the time-lapse of blood degradation [37]. Most often, these lesions express a "ground

#### *Endometriosis - Recent Advances, New Perspectives and Treatments*


**Table 4.**

*Studies that evaluated the diagnostic utility of the "ground glass" appearance of endometriomas.*

glass" appearance, a feature to which different degrees of accuracy have been attributed over time (**Table 4**).

In practice, the grayscale US of endometriomas and HCs can be very similar, both lesions showing characteristics of different stages of blood degradation, making the distinction difficult [42]. A study [43] showed that 20 texture parameters that were extracted from the US of greyscale images showed statistically significant results when distinguishing endometriomas from HCs. Their combined ability was able to differentiate between the two entities with 100% (95% CI, 88.4–100%) sensitivity and 100% (95% CI, 75.3–100%) [43]. Three parameters were proved to be independent predictors of endometriomas (difference variance, contrast, and the 10th percentile) (**Figure 6**) [43]. The difference of variance parameter measures the variance of the difference of gray level values (reflecting the randomness within an image) [44]. The contrast parameter shows the local variations present in an image, expressing higher values when an image contains a large number of pixels with different gray level values [45]. The study [43] showed that both of these parameters held higher values for HCs than for endometriomas. On the other hand, the 10th percentile showed higher values for endometriomas than for HCs, which signifies that 10% of the pixels within images were distributed under higher values for endometriomas than for HOCs [43]. Even though endometriomas were expected to have a higher degree of echogenic randomness due to a large number of contained elements, HCs displayed higher values for the parameters that mirror these characteristics [43]. This finding is consistent with the literature, which suggests

#### **Figure 6.**

*(A) T2-weighted image of a patient with histologically-proven endometrioma. The texture maps show the distribution of the (B) 10th percentile, (C) contrast, and (D) difference variance parameters.*

that HOCs have more complicated and heterogeneous content on TVUS (since they express tiny linear strands and retraction clots more frequently) [46].
