Automatic segmentation of the spinal cord from MR scans and differential diagnosis of MS lesions with deep learning architectures

The human spinal cord is a highly organised and complex part of the central nervous system and its function is to transmit neural signals from the brain (sensory information) to the peripheral nervous system (motor information) and from the peripheral nervous system to the brain. When MS (Multiple Sclerosis) occurs in the spinal cord, it affects the white and grey matter of the brain, the spinal cord and the optic nerve. Early diagnosis of MS is important to slow the progression of the disease and control symptoms. Starting the right treatment early can prevent the disease from causing more severe attacks and improve the patient's quality of life. In this way, it may be possible to stop or slow the progression of MS. Clinical symptoms/signs, cerebrospinal fluid tests, evoked potentials and Magnetic Resonance Imaging (MRI) findings are used to diagnose MS. In particular, the widespread use of MRI and the development of computer-aided systems have contributed significantly to the diagnosis and follow-up of MS. On the other hand, studies based on the segmentation of the spinal cord from MR images using artificial intelligence algorithms, especially deep learning models, and the presence or absence of MS lesions in the spinal cord region have also become prominent in recent years. However, although these and similar studies have achieved a certain level of success, it can be seen that the success in MS detection in these studies is low due to reasons such as the small amount of data due to the small size of the dataset and the small volume of MS lesions. In this thesis, the segmentation of cervical spinal cord cross-sectional area (CSA) and cerebrospinal fluid (CSF) areas on T2-weighted MR images taken from different planes such as axial and sagittal with deep learning and differential diagnosis of MS lesions in the spinal cord were performed. In the study, a dataset was first prepared to perform segmentation of cervical spinal cord CSA, CSF area and MS lesions within the spinal cord boundaries using cervical spinal cord MR data obtained from Akdeniz University Hospital. In this dataset, FractalSpiNet, Con-FractalSpiNet and Att-FractalSpiNet architectures, developed based on U-Net architecture, were used to segment the spinal cord and CSF areas in MR images in sagittal and axial planes, and to detect MS lesions within the spinal cord boundaries. In addition, the results obtained with the proposed architectures are also compared with mixed architectures, namely Att U-Net (Attention U-Net), Res U-Net (Residual U-Net) and Att-Res U-Net (Attention Residual U-Net). In this thesis, spinal cord axial CSA/CSF, spinal cord axial MS and spinal cord sagittal MS data subgroups were created to mask the areas to be segmented in the cervical spinal cord dataset. For segmentation of the spinal cord area and detection of MS lesions on the prepared cervical spinal cord dataset, DSC (Dice Similarity Coefficient) based on pixel similarity was used to measure model success, PRE (Precision) and REC (Recall) metrics were used, VOE (Volumetric Overlap Error) and RVD (Relative Volume Difference) as volume-based metrics, and ASD (Average Surface Distance) and HD95 (95th percentile Hausdorff Distance) as distance-based metrics. Firstly, experimental studies were performed on the axial CSA/CSF sub-dataset of the spinal cord. At the end of the model training, the best results were obtained with 94.99% DSC score with Con-FractalSpiNet architecture for CSA segmentation, 93.00% DSC score with FractalSpiNet architecture for CSF region and 96.54% DSC score with FractalSpiNet for segmentation of the whole spinal cord region. As a result of the training performed on the other subset of MS spinal axial data, the best results for the first segmentation region, the CSA, were obtained with the Con-FractalSpiNet and FractalSpiNet architectures with DSC scores of 98.89% and 98.88% respectively, while the best results for MS lesion detection were obtained with the Con-FractalSpiNet and FractalSpiNet architectures with DSC scores of 91.48% and 90.90% respectively. In the same data subset, the most successful models for segmentation of the non-MS spinal cord area were Con-FractalSpiNet and FractalSpiNet with DSC scores of 97.25% and 97.17%, respectively. When analysing the experimental results on the sagittal spinal cord MS data subset, 97.06% and 95.16% DSC scores were obtained with the Att-Res U-Net architecture as a result of model training for segmentation of the spinal cord area and spinal cord areas without MS, while the most successful results were obtained with a DSC score of 56.25% using Con-FractalSpiNet for detection of MS lesions. When all the results are evaluated, using the proposed U-net based FractalSpiNet architectures, highly competitive results were obtained in the segmentation of the cervical spinal cord region and MS lesions in this region compared to existing studies.