Analysis of deep learning methods using magnetic resonance images for the diagnosis of Multiple Sclerosis disease

Multiple Sclerosis (MS) is a chronic neurological disease that frequently affects young and middle-aged individuals and impacts the central nervous system (CNS). This disease, which causes inflammation in the brain and spinal cord, leads to damage to the myelin sheath surrounding nerve cells, resulting in disruptions in signal transmission. This can lead to cognitive function losses and functional disabilities. In the later stages of the disease, irreversible nerve damage can occur. These conditions in MS patients make it extremely important to diagnose the disease early and closely monitor its progression. The most commonly used imaging method for MS diagnosis is Magnetic Resonance Imaging (MRI), and the diagnosis is generally made by observing MS lesions in scans performed at two different time intervals. Monitoring the progression of the disease is done by tracking the time-dependent changes of these lesions in MRI slices. However, the small size of MS lesions and their similarity to other neurological diseases make accurate diagnosis difficult. Since MS is characterized by an inflammatory process in different areas of the brain, different symptoms emerge in each patient, making the diagnostic process challenging. MRI is one of the most important neuroimaging techniques for observing structural changes in the brain and spinal cord in MS diagnosis. MS plaques in the brain can be clearly visible with MRI. The activity of lesions can be evaluated using gadolinium (Gd), a paramagnetic substance administered through the bloodstream; active, or acute MS plaques retain gadolinium, while chronic plaques do not absorb this substance. Therefore, gadolinium uptake indicates that a lesion is active. In recent years, the development of deep learning (DL) and artificial intelligence (YZ) techniques has made significant contributions to the analysis of MRI images. Deep learning algorithms, which automatically learn from large datasets, perform well in tasks such as feature extraction, classification, and segmentation, and are used as critical auxiliary tools in the diagnosis of MS and similar neurological diseases. For example, the ResNet101 model reduces overfitting in deep neural networks, allowing for more efficient learning, while networks such as VGG16 and DenseNet121 provide high accuracy in image classification. Especially, DenseNet121 stands out with its faster runtime time and high accuracy rates, while InceptionV3 can perform more detailed feature extraction using multi-dimensional filters. MobileNet, on the other hand, is optimized for mobile and embedded devices, providing high performance with low hardware resources. In this thesis, a new DL-based YZ model for analyzing MRI images for MS diagnosis has been developed. This model provides more feature extraction from pre-trained network architectures through a Feature Integration method. Low-level and high-level features are combined to obtain the final feature map. Features extracted using pre-trained VGG16, ResNet101, InceptionV3, MobileNet, and DenseNet121 networks are combined with deep convolutional layers. The dataset used in the experiments consists of axial and sagittal FLAIR brain MRI images prospectively obtinated from 72 MS patients and 59 healthy male and female individuals who applied to the Turgut Özal University Faculty of Medicine in 2021. Three subsets of the dataset were created for binary classification between MS and healthy: axial images only (n = 1652), sagittal images only (n = 1775), and combined axial and sagittal images (n = 3427). DenseNet121 achieved an accuracy of 98.78% for axial images, while the VGG16 model achieved 99.32% accuracy for sagittal images. Moreover, the FPS (Frames Per Second) performances of the developed models were compared with similar models in the literature. The ResNet101-based model showed a slight performance improvement with 28 FPS compared to the literature value of 27 FPS. The DenseNet121-based model showed a more significant improvement with 37 FPS, compared to the literature value of 32 FPS. The InceptionV3-based model demonstrated higher performance with 39 FPS compared to the literature value of 33 FPS. However, in the VGG16-based model, a reverse situation was observed; the proposed model performed at 87 FPS, falling behind the literature value of 123 FPS. This performance drop is attributed to feature extraction from intermediate layers in the proposed model, where the traditional hierarchical convolutional structure of VGG16 decreased its efficiency in this process. On the other hand, innovative convolution blocks like multi-dimensional filters in InceptionV3 and skip connections in ResNet101 prevented performance loss in FPS and even led to an increase. Overall, most of the proposed models provided higher FPS values compared to models in the literature, resulting in significant improvements in processing speed. These results demonstrate that deep learning-based systems can be used as an effective decision support tool in MS diagnosis. The development of such Computer-Aided Diagnosis (CAD) systems will help physicians in early diagnosis and contribute to preventing the disease from negatively impacting individuals' quality of life.