Segmentation of breast microwave imaging using fuzzy c-mean clustering

Abstract

Breast cancer is the uncontrolled growth of breast cells, which grows mainly in fatty tissue and lobules. It could also spread over the blood or lymph system to other organs, causing virulent metastasis. In 2020, Female Breast cancer was the most frequently diagnosed cancer with 2.3 million cases, based on World Health Organization's (WHO) reports. According to statistics, 684 996 deaths from the disease were recorded in the same year. The most effective way to control the spread of breast cancer and reduce the number of deaths is early detection and treatment through clinical evaluation. There are several imaging techniques for screening female breast tissue abnormalities and cancer, such as Magnetic Resonance Imaging (MRI), Ultrasonography, Positron Emission Tomography (PET), and Mammography. Currently, mammography is considered to be the gold standard methodology in order to screen breast cancer. However, in accordance with the mammography restrictions, such as high radiation dose and low sensitivity in dense breasts, the mentioned method is not remarkably efficient. For this purpose, other screening methods are necessary to eliminate these restrictions. Microwave imaging (MWI) is a new modality that could potentially become a supplementary method, to already existing ones, to contribute the breast cancer diagnosis at an early stage, due to using non-ionizing radiation, safe and inexpensive technology. MWI is based on differences between the dielectric properties of cancerous and normal breast tissue. Moreover, lesion segmentation is essential for diagnosing breast cancer. To do this, specialists examine the images manually, which may lead to misdiagnosis due to human visual perception error. Also, in some cases, this process can be time-consuming. In order to solve these problems, automated segmentation methods have been utilized to reduce these errors. Clustering algorithms or unsupervised machine learning techniques could be employed to segment images to minimize human involvement in segmentation and achieve a precise detection of the cancerous lesion. In this thesis, a segmentation method is suggested to extract lesions from low-resolution MWI images with the help of the Fuzzy C-mean algorithm and statistical features. Microwave images used in the thesis were taken from a clinical study. The study's codes are all implemented in MATLAB. To evaluate the effectiveness of texture features of MWI for segmentation by FCM algorithm, three statistical features in the spatial domain i.e., intensity, entropy, and energy were used to form the feature vector. Two series of experiments were performed on the data. The first experiment included a combination of features consisting of intensity, energy, and entropy matrix of each image were form. Energy and entropy matrices were calculated for each image and combined with intensity matrices and applied to FCM. No difference was observed regarding the effect of entropy and energy matrices in segmentation with only intensity features. In the second experiment, only different combinations of intensity (i.e. color features) of the images were used. Three feature matrices were formed using red, green, and blue planes of each image, in RGB color space. A three-dimensional matrix using RGB planes, a matrix using the mean of RGB planes, and at the end, only red plane as feature matrix, were formed and applied on FCM algorithm as the first trial, separately for each image. The segmented images compare to ground truth images. The experimental results demonstrate that the segmentation method using only the red plane, to segment the dark red region of the image, which represents the highest values of dielectric activity of the tumor, outperforms the other two methods in the case of most overlapping with the ground-truth image. Also, in respect of specificity, this method has performed better compared to multi-channel methods. However, in terms of sensitivity, the red single-channel method does not provide the expected reliable results. Finally, the achieved size of the segmented tumor in each trial was compared with the size of the tumor in conventional imaging measured by a specialist. In certain sizes, there are acceptable achievements, but for satisfying results, existing algorithms need more improvements.