Follow-up and automatic detection of multiple sclerosis lesions using deep learning models on MR scans

Multiple Sclerosis (MS) is a common central nervous system (CNS) disorder that affects the brain and spinal cord, which is common in young and middle-aged people. The disease that occurs as a result of inflammation in the sheath structure on the neural networks in the CNS causes significant cognitive losses in the person. Loss of functional ability and irreversible brain damage can be seen in advanced stages in MS patients. For these reasons, early detection and follow-up of MS disease is very important. Magnetic resonance (MR) imaging is common in the diagnosis and follow-up of MS, and MS lesions are generally expected to be seen on scans of at least two different periods. The follow-up of the disease is determined by following the changes of the lesions in the MR sections. MS lesions are small in size and resemble other neurological disorders in the brain, making it difficult to detect MS lesions. In this thesis, an improved deep learning model (iMask R-CNN) is proposed for automatic detection of MS lesions on MR images. In this model, the prediction important region alignment (RoIAlign) process of Mask-Based Regional Convolutional Neural Network (Mask R-CNN) architecture is improved. In addition, an MR dataset named MSAkdeniz has been prepared, which includes FLAIR MR sections periodically obtained from 52 patients for the detection of MS lesions. Moreover, a web-based decision support system (DSS) application named DeepMSWeb, which uses the iMask R-CNN model, has been developed to help physicians for automatic detection of MS lesions and monitoring of change. For the automatic detection of MS lesions, experimental studies are carried out on MSAkdeniz dataset and eHealth, UMCL, ISBI2015 and MICCAI2008 datasets on two different platforms, using five different deep learning models, primarily the proposed iMask R-CNN model, and the results are compared. Using the proposed iMask R-CNN, scores of 88.32±4.44% for Dice similarity coefficient (DSC) and 89.80±3.54% for lesion true positive rate (LTPR) are obtained in lesion detection datasets. With the proposed iMask R-CNN and DeepMSWeb, assistant tools that can be used in the detection and monitoring of MS lesions have been developed.