Unet based segmentation in qualitative microwave imaging for breast cancer diagnosis

Abstract

Breast cancer is a prevalent type of disease which affects millions of people every year. Early diagnosis of this disease is critical to save the patients' lives. The traditional diagnosis methods of breast cancer have some drawbacks, such as including ionising radiation, low resolution, age limitations, and a long imaging procedure. To overcome these drawbacks, researchers have developed the alternative imaging methods, such as microwave imaging. Microwave imaging is based on the using electromagnetic waves, therefore this imaging procedure is safer than other imaging techniques which include ionising radiation. Microwave images have low resolution because of the generation of the imaging process, the inverse scattering problem occurs. This problem affects the image quality in terms of the examination of biomedical imaging. To overcome this issue, researchers utilize qualitative or quantitative imaging algorithms. However, the microwave breast cancer images are required to evaluate by a radiologist. This thesis has two main observations. In first part of the thesis, the microwave breast cancer images will generate by using a qualitative imaging method- the factorization method, to obtain the images from the total electric field. Another observation of this thesis is the detection of whether the tumour region exists with the help of a segmentation process by using deep learning methods. This thesis aims to present a solution to the detection of breast cancer in real time with the help of the properties of deep learning methods. Thus, it will be presented a contribution of the literature as a comparison of the model performance results of this study. The synthetic microwave breast cancer images were generated by using the factorization method, and the dielectric profiles of the images were determined based on the ex vivo studies of the microwave breast imaging. The 1500 images were obtained from this process. This dataset contains the 1000 images of single tumours, the 250 images of two-tumourous and the 250 images of without tumour. The ground-truth images were generated based on the dielectric profile of the tumorous regions. The Unet, Attention Unet and Attention Residual Unet models were used for segmentation of the images. The optimum hyperparameter values were determined by using Optuna.The pixel-based image similarity metrics, such as SSIM and MSE, were used to evaluate these models results. As a result of this thesis, the use of the Unet model and other models showed significant image segmentation performance by using the synthetic image dataset obtained from the total field. The mean of the SSIM values of the Unet model was found to be 0.94, the Attention Unet was 0.9278, and the Attention Residual Unet model was 0.9236. The mean of MSE value of Unet model was found as 0.00973, the Attention Unet was 0.01609 and the Attention Residual Unet model was 0.01393. These models performed the segmentation process in seconds. The training and test duration of the Unet model was shorter than other models due to the configuration of the model. This thesis is proposed as a preliminary study of the detection and segmentation of the suspicious tumour regions due to we have known the properties of the dielectric profile of the examined tumorous regions. In reality, the localisation and size of tumour regions are required to label by a radiologist for use in segmentation tasks. The lack of this study, it was proposed to use the two-dimensional slice microwave images of the breast, but the nature of the examined organ is three-dimensional. Therefore, the three dimensional qualitative imaging methods and the three-dimensional deep learning methods for segmentation of the tumorous regions could be utilized.