Automatic detection of white matter hyperintensities using deep learning techniques on brain magnetic resonance images

Detection and correct classification of abnormalities in the white matter part of the brain are of vital importance in terms of minimizing the physical and cognitive damage that may occur by early diagnosis of diseases. Abnormalities called white matter hyperintensity (WMH) can be observed even in healthy individuals who have magnetic resonance images (MRI) due to any complaint, and physicians have serious difficulties in deciding whether or not WMH formations are harmful. In this thesis, an automated approach, realized with high-performance deep learning networks, has been presented to facilitate physicians' work for WMH detection and segmentation by automatically scanning magnetic resonance (MR) images. The study consists of two parts. In the first part, it is revealed that instance segmentation can be used for WMH segmentation, which is an advanced segmentation method that can contribute more to the decision-making processes of physicians, but is not widely used in difficult problems such as WMH segmentation. As a method, Mask-Based Regional Convolutional Neural Network (Mask R-CNN), which was developed for automatic segmentation of WMHs with instance segmentation and contains many hyper-parameters, was used with fine tuning. Three different datasets were used, one of which was the tumor dataset created specifically for this thesis study, and the other two were open-source tumor and multiple sclerosis (MS) imaging datasets. Two WMH classes were trained separately and together, in the form of brain tumor and MS datasets. As a result of the tests performed with the model obtained by training all the data together, WMHs in the form of MS lesions and brain tumors in MR sections were successfully detected with a high mAP score of 0.94 according to the mean precision (mAP) metric of all classes. In addition, according to the Precision ratio (PRC) and Dice similarity coefficient (DSC), segmentation performance of 0.86 and 0.82, respectively, was obtained. In the second part, two different deep learning methods, U-Net and Mask RCNN, were used for comparative analysis. The larger datasets used and the application of data augmentation approaches that yield more effective results have increased the performance and generalization ability of the model, and this has been demonstrated by appropriate measurement metrics and loss functions. When the results are compared with the studies in the literature, with instance segmentation, a very close and relatively better performance was obtained with the semantic segmentation result. In addition to DSC and PRC, the sensitivity (RC) metric and the F1 metric, which is the harmonic mean of precision-sensitivity values, were used for performance evaluation. For the stroke dataset; performans values of 0.93 DSC, 0.97 PRC, 0.98 RC and 0.98 F1 performances were achieved with Mask R-CNN, 0.92 DSC, 0.89 PRC, 0.95 RC and 0.92 F1 performances were achieved with U-Net. For the WMH dataset; performance values of 0.83 DSC, 0.83 PRC, 0.73 RC and 0.78 F1 with Mask R-CNN; performance values of 0.82 DSC, 0.82 PRC, 0.83 RC and 0.81 F1 were achieved with U-Net. In addition, for different network configurations, training-test times and performance results were compared and analyzed extensively. When the experimental studies conducted within the scope of the thesis study and the results of the studies on WMH segmentation in the literature are compared, it was concluded that the Mask R-CNN method can achieve higher performance in WMH segmentation and detection, and the performance rate can be further increased with the use of larger dataset and more powerful hardware.