Machine learning based augmentation of medical microwave imaging

Abstract

The main reason why microwave imaging, which is an alternative to in-body imaging methods, is desired to be used in biological systems is that it is not ionized. Other reasons why it is desired to be used for the detection of udder tumor can be said that udder compression is not performed and the device materials are more affordable. Each tissue has its dielectric properties different from another tissue. However, if the tissue has the same tissue but is tumorous, it will now have different dielectric properties due to the excess blood flow in it. Quantitative methods, one of the classes of microwave inverse problems, can also show the permeability and conductivity properties of the scatterer, however, it has been stated that only position and shape information can be determined in the qualitative methods, which are stated as computationally efficient and fast in the literature. However, with the solution of qualitative inverse problems, apart from the position and shape information of the scatterer, the electromagnetic properties of the scatterer should also be determined. In addition to the well-known qualitative inverse methods in the literature, linear sampling method (LSM) and, factorization method (FM), inverse time migration (RTM) method is also used to reconstruct the images of scatterers. If the pixel changes in the images created using these methods are examined and correlated with the dielectric properties of the scatterer, this threshold of qualitative inverse problems can be eliminated. In the first part of the thesis, microwave images were created with qualitative inverse methods (LSM, FM, RTM) using s-parameters taken from a scatterer in the simulation environment. To capture the relationship between the images and the dielectric properties of the scatterer, graphs were created by calculating the maximum pixel values and averages of the location region of the object in the microwave images. When these graphs are examined, it is observed that there is a general logarithmic similarity between the pixel values in the microwave images despite the changing dielectric and position information of the scatterer. Simultaneously with this process, comparisons were made between the qualitative reverse imaging methods and the desired synthetic images by looking at the similarity of MSE values. In addition, the factors affecting the image were also observed for the scenario determined while creating the images. Thanks to the relationship between the scattering microwave images and their dielectric properties, machine learning methods were used to make classification and prediction. A dataset was created with images in different locations and dielectric properties to be used in machine learning. These images are simple microwave images and it is aimed to create more meaningful results with only the key information of the images by using the SVD method. As the machine learning algorithm used here, Adaboost, which is considered to be suitable for the data set features, was used. In the first stage, a dataset containing only microwave image singular values and dielectric permittivity values was prepared and classification was made. When the confusion matrix formed as a result of the classification was examined, it was observed that a high-accuracy classification was made. In the second stage, the possibility of predicting the singular values of the ground truth images with the singular values of the microwave images was examined. The data set at this stage consists of synthetic image and microwave image singular values with the most significant weight. As a result, results with high predictive MSE values were obtained. The images were reconstructed using the synthetic image singular value estimates obtained from these results and the real microwave image SVD value. The similarity correlation coefficients of the images were calculated by comparing the predicted image and the microwave image with the desired synthetic image. The estimated image pixel values were magnified and brought closer to the desired dielectric value. The second stage of the thesis is aimed to design a breast phantom similar to the real breast tissue in the simulation environment and to make a dielectric estimation with the images obtained from it. Similarly, a new dataset was created from the breast phantom data. Similar to the processes performed during the previous machine learning processes, a breast phantom microwave image synthetic image and dielectric profile not included in the dataset were estimated. By replacing the real microwave image SVD values with the singular values found, the image pixel value was moved to higher values and approached the synthetic image values. The estimated image and the microwave image obtained as a result of breast phantom simulation were compared with the desired synthetic image by looking at the similarity values. As a result, it can be said that the predicted image improves the microwave image.