Diagnosis of brain cancer and contour normal tissue for radiation therapy based on deep learning methods

Abstract

Brain tumors are one of the deadliest types of cancer ever identified. Rapid and accurate diagnosis of brain tumors, followed by surgical intervention or appropriate treatment, increases the probability of survival. Accurate identification of brain tumors in MRI scans allows precise location of surgical intervention or chemotherapy. Accurate segmentation of brain tumors in MRI scans is challenging due to their varied shapes and requires knowledge and accurate image interpretation. This thesis starts with analyzing machine learning and traditional methods and focuses on the study of edge detection using the Sobel and Canny edge detector algorithm. After that, we use morphology-based techniques to segment the images and evaluate the results. We use K-means techniques for Clustering. Despite various advances, these methods still show limitations in complex situations such as tumor detection and segmentation. In the next step, we analyze the process of dividing photos into parts using transformations. Specifically, we discuss the Wavelet and Contourlet transforms. By using these transformations, we get more detailed information about the analysis of the images. These transformations have many applications, and we can identify the borders of the image and combine them. Finally, we can use this transformation to process and generate deep learning masks using a supervised model. In the following, we analyze new techniques using supervised and deep learning approaches in two specific areas: image classification and image segmentation. As we introduce these methods, we introduce the obstacles facing deep learning and discuss potential strategies to solve and enhance them. Using deep neural networks and the Resnet 50 model, we classify brain images into tumor and non-tumor categories and achieve a satisfactory score of 97% in the F1 criterion. In addition, we introduce and analyze the Unet deep network in deep learning and upgrade it to a RESUNET network for segmentation. The results of this segmentation show that the proposed approach, with different criteria, such as the DICE metric with a score of 0.9434, performs exceptionally during training compared to conventional topologies and shows a faster convergence rate. In the last part, we presented the unsupervised learning system and developed the adversarial generative network to generate brain MRI images. The adversarial generative network is an intelligent network for generating the desired data, and the results show the effectiveness of the adversarial generative network in generating new data. It is of exceptional quality.