Anomaly detection in brain MRI images with deep learning methods

Fast and accurate detection of brain lesions is important for treatment, and automatic detection of these lesions is a research topic, and computer-aided systems are being developed for anomaly detection from brain MRIs. Current solutions and technological methods rely heavily on the development of supervised learning models detection methods using labeled MRI data, low performance, lack of data sets or diversity make it difficult to develop successful applications. It is important to develop unsupervised approaches to eliminate these and therefore facilitate clinical use, and to develop unsupervised learning-based approaches to partially eliminate or facilitate the difficulty of obtaining labeled MRI data. Developments in deep learning methods, modeling of high-dimensional data with deep web-based unattended learning methods, strong analysis and high performance examples and applications have paved the way for developing alternative solutions in different fields. In this thesis study, in order to detect lesions with non-consultative learning models, new detection models based on Producer Controversial Network (GAN) that learn normal brain MRI images and healthy brain image patterns and then differentiate tumor brain images are proposed. The proposed GAN model was trained with the healthy patient MRI data, and lesion detection was made from the outputs produced as a result of outliers that did not fit the model. The GAN model is based on the assumption that suitable occult area representation cannot be found for lesioned MRI images, as it is trained with healthy MRI data. The hypothesis is that reconstructed MRI images are expected to give higher reconstruction error than healthy MRI images. Within the scope of the thesis, three different new solution proposals were presented, GAN was applied to tumor detection for the first time, and improvements were made by modifying GAN models. First, a new GAN-based inverse mapping model is proposed that enables the transition from image space to hidden space. In the second model, MRI section information was input to the model to increase the success of reverse mapping using the Conditional GAN architecture and successful results were obtained. In the last model, the success of the model is increased by using the Wasserstein distance, which contributes to the stable training of the GAN model. The proposed models have been analyzed and tested with two well-known data sets. The results obtained show that the proposed models will be successfully applied not only in tumor detection but also in other areas.