Volumetric quantification in ovarian pathology using abdomino-pelvic computed tomography

Pathological ovary is categorized into cystic tumors, solid tumors and mixed, according to the content of the affected ovary. Accordingly, the degree of benignity or malignity thereof is established. The imaging study for the preliminary morphological assessment of PO is ultrasound, in its pelvic and transvaginal modalities, for which well-established criteria are available. Once the ultrasound findings suggest malignancy, complementary studies such as abdominal-pelvic tomography images and tumor markers are requested. This type of images has challenging problems called noise, artifacts and low contrast. In this paper, in order to address these problems, a computational technique is proposed to characterize a pathological ovary. To do this, a thresholding and the median and gradient magnitude filters are applied, preliminarily, to complete the preprocessing stage. Then, during the segmentation, the algorithm of region growing is used to extract the three-dimensional morphology of the pathological ovary. Using this morphology, the volume of the pathological ovary is calculated and it allows selecting the surgical-medical behavior to approach this kind of ovary. The validation of the proposed technique indicates that the results are promising. This technique can be useful in the detection and monitoring the diseases linked to pathological ovary.


Introduction
The ovaries are two nodular bodies located one on either side of the uterus, attached to the broad ligament, and situated close to the lateral pelvic wall in the ovarian fossa [1].
On the other hand, the survival rate of epithelial ovarian cancer (EOC) is approximately 35%-40% despite maximal treatment efforts, highlighting a need for stratification biomarkers for personalized treatment [2].
The modalities of medical imaging play a crucial role in the diagnosis of ovary lesions. Multilayer computed tomography (MSCT) and ultrasound technique are is the primaries diagnostic methods for ovary tumors. Additionally, the segmentation of ovary tumors is a problem of great interest because the three-dimensional (3D) segmentation of these tumors from MSCT images is a prerequisite for computer-assisted diagnosis, treatment planning and cancer control of ovary. The main imperfection of MSCT images is the stair-step artifact, which dramatically deteriorates the appearance of such objects and can affect the dimensions especially of small structures present in the images acquired  [4]. This is a real problem due the ovaries are a very small anatomical structures and this type of artifact can affect severely the associated information with the ovary.
Several works related to the segmentation of ovary or ovary tumor, which are presented at next. In this sense, Cigale and Zazula performing the ovarian segmentation in ultrasound images using cellular neural networks obtaining an accuracy of 70% [3]. Ramya and Kiruthika propose a technique to detect ovary follicles present in ultrasound images. For this, they incorporate discrete wavelet transform based k-means clustering and they report an accuracy of 96.11% based on structural similarity method [4]. Sonigo et al. show a strategy based on deep learning to detect and to count ovary follicles in mouse ovary and they affirm that this strategy has an accuracy of 91.36% [5].
In this paper, the main purpose is to generate a semi-automatic technique (SAT) for pathological ovary (PO) segmentation, in MSCT images. This technique is based on the application of a filter bank (thresholding, median and gradient magnitude filters) and region growing technique in order to generate the 3D PO morphology. This bank is necessary for address the problems of the aforementioned imperfections of MSCT images. This technique can be useful in the detection and monitoring PO for establishing the main actions oriented to address the pathologies associated with PO.

Dataset
A three-dimensional MSCT database was used, which has a voxel number of 512 x 512 x 70. Manual segmentation of the ovary by a gynecologist is also available.

Computational strategy proposed
In Figure 1, a schematic diagram is presented. It synthesizes the computational algorithms that make up the SAT for the segmentation of the human ovary. In this figure, the MSCT dataset matches with the dataset described in the 2.1 section of this paper. The other elements, shown in Figure   a) Thresholding filter. Two thresholds linked to ovary detection were considered to isolate ovary in MSCT datasets. The thresholding method, used in this article, considers a histogram for threshold selection [6]. b) Median filter. A median image (MI) is generated processing each original image (OI) with a median filter [7]. The role of this filter is to address the noise present in the images. The tuning parameter of this filter is the size of the neighbourhood [8]. c) Gradient magnitude filter (GMF). In this work, an approach based on finite differences was used for GMF computational implementation [9]. This filter generates a smoothed version, called GM $ , calculating the three-dimensional gradient magnitude of MI, using the mathematical model given by Equation (1). (1)  (1), is very expensive. By this reason an approach based on central finite differences is used, in this paper, for modelling computationally the GMF [10].

Segmentation.
In this stage the region growing technique is applied. Region growing technique (RG). The RG partitions an image (f) into regions (Ri) whose voxels are connected according to certain predefined criteria based on connectivity and similarity of the image. The most popular predefined criterion is given by Equation (2) [10].
6f(i, j, k) − µ ? @ 6 < mσ ? @ The RG needs a seed voxels into an initial neighborhood (Iv) being: f (i, j, k) the gray levels of Iv, µ ? @ the average gray levels of a Iv of arbitrary shape and size, k an arbitrary scalar and σ ? @ the standard deviation of an arbitrary neighborhood of Iv. The Equation (2) represents the most popular predefined criterion due its simplicity and effectiveness.
In this point, it is necessary to notice that the RG tuning parameters are the initial neighborhood size (r) and k parameter that controls the amplitude of the range of intensities considered to accept or reject a voxel in a region. Such parameters must undergo a tuning process. During the tuning process, the PO segmented is compared with the manual segmentation traced by a gynecologist. The Ds is used in order to estimate the difference or matching between these structures [12].

Results
First, we present qualitative results using original and preprocessed images in two-dimensional (2D) and 3D views. Then the quantitative results are shown considering mainly the results accuracy and optimal parameters obtained by the algorithms present in the proposed SAT. Figure 2 shows a 2D view of both the original image and the processed images after applying the proposed technique to the dataset considered.   Figure 2, we can observe the effects generated by filter bank application. A qualitative analysis revels that the considered filters generated high quality images reducing the noise and producing a high contrast between the anatomical structures. In this figure, we can observe the excellent performance of gradient magnitude filter for generating the contours of aforementioned structures. The threshold filter eliminates non-important anatomical structures and gives an adequate split image (see Figure 2(b)). The median filter generates an almost-free noise image that facilities the PO segmentation (see Figure 2(c)); whereas the Figure 3(b) shows the enhancement contours linked to the pathological ovarian.

Qualitative results
On the other hand, Figure 3 illustrates the results of the segmentation process developed using region growing technique. In this figure, we can observe an excellent representation of segmented kidney stone morphology in both 2D and 3D views.

Quantitative results
When the tuning process was performed the PO was characterized considering its volume. So, the volume occupied by the semi-automatic segmented PO was 8.41 cm 3 ; while the volume, reported by the clinical expert, (obtained considering the manual segmentation) was 8.59 cm 3 . The percentage relative error, considering these volumes, was 12.1%. This error represents an accuracy of 87.9%. Additionally, the maximum Dice coefficient generated was 0.82. This result allowed establishing the optimal parameters, of the computational algorithms that make up the proposed technique, which are presented below: a) Median filter: the size of the kernel was corresponding with (3 x 3 x 3). b) Region growing: r = 1 and k = 4.50. Additionally, the Table 1 shows comparative information about the accuracy obtained both in this paper and in others researches, reported in the specialized literature.

Authors
Technique Accuracy (%) Cigale, et al. [3] Cellular neural networks 70.00 Ramya, et al. [4] k-means and wavelet transform 96.11 Sonigo, et al. [5] Deep learning 91.36 Valbuena et al. (Our technique) SAT 87.90 The information analysis presented in Table 1, let us to affirm that SAT was outperform by the results reported by Ramya, et al. [4] and Sonigo, et al. [5]; whereas the computational technique propose in this paper generated a higher accuracy value than that reported by Cigale, et al. [3]. It is important to notice that the techniques mentioned in Table 1 the reported in [5] is based on smart

Conclusions
A semi-automatic technique, available for detecting pathological ovary in a precise and efficient manner, has been presented. This technique was based in the application of thresholding, median and gradient magnitude filters in order to address the noise problem; whereas the region growing technique was applied for generating the 3D PO morphology, presents in MSCT images.
The three-dimensional representation of PO is useful for the detection and monitoring of ovaries diseases; as well as for the planning of medical treatments and clinical-surgical procedures linked to pathologies associated with the human ovaries.
On the other hand, it is expected that the segmentation generated by the proposed method can be useful to promote, deepen and potentiate the study of the real anatomy of the structures linked to the ovaries. In the immediate future it is planned to validate the proposed technique with a significant number of databases in order to estimate the robustness of the aforementioned technique.