Computer aided diagnosis of intervertebral disc pathology in lumbar spine
Alomari, Raja' S.
MetadataShow full item record
Lower back pain (LBP) is the second most common neurological ailment in the United States after the headache according to the National Institute of Neurological Disorders and Stroke (NINDS) . Americans spend at least $50 billion each year on health care of people with LBP. There are over 11 million Americans who have some sort of Intervertebral Disc Disease (IDD) . Intervertebral disc degeneration (e.g., desiccation and herniation) in the lumbar area is one of the most common diseases that cause LBP and sciatica, a common term for pain in legs consequent to irritation of the sciatic nerve [9, 10]. Neuroradiologists use clinical Magnetic Resonance Imaging (MRI) for diagnosis of most of the lumbar area abnormalities and specifically the disc degenerative diseases. Clinical MRI usually comprises sagittal and axial T1- and T2-weighted co-registered protocols beside some other protocols to help the radiologists in decision-making. They diagnose hundreds of cases each day that range from simple diseases to major life threatening diseases such as tumors. Thus, full or partial automation of the diagnosis process reduces the large burden on them and increases their effectiveness. Besides that, introducing reliable and robust full or partial diagnosis system reduces the inter-observer variability and increases the degree of confidence of the diagnosis. There are two main steps that the neuroradiologists follow during the work flow of diagnosis of lumbar area: (1) Detection and labeling of the discs (and /or the vertebrae), and (2) Diagnosis of each disc (and vertebra) level by specifying all possible abnormalities in that disc (and vertebra). The detection and labeling step is required only for one of the sagittal views because all provided protocols (T1- and T2-weighted for sagittal and axial views) are manually co-registered by the technician during the acquisition of the MRI. However, the neuroradiologists identify all possible abnormalities at each disc level. They also identify vertebral bone disorders if there is a potential vertebral disorder and might recommend Computed Tomography (CT) scan for better visualization of the bone. In diagnosis, there are many pieces of information that the neuroradiologists utilize before making the decision such as age and height of the patient, history and type of pain. They also utilize the relative context of the discs and the vertebrae. In this dissertation, we propose machine learning-based solutions for diagnosis of lumbar intervertebral disc degenerative diseases (IDD). We study clinical MRI from our collaborating neuroradiologists group and use their clinical diagnosis reports as our gold standard of the ground truth for diagnosis of each disc level in the lumbar area. We, initially, perform robust and automatic detection and labeling of the lumbar intervertebral discs from sagittal views. We develop a probabilistic model for localization of the six discs in sagittal T2-weighted MRI for the lumbar area. In our model, we incorporate two levels of information: low- and high-level. In the low-level, we model the local pixel properties of discs, such as appearance. In the high-level, we capture the object-level geometrical and contextual relationships between discs. We estimate the model parameters from manually labeled cases (supervised learning). Then we perform machine-learning based analysis at each localized disc for further diagnosis tasks. Our proposed model investigates degenerative abnormalities and allows the flexibility of adding features depending on the abnormality. It also allows incorporation of necessary domain knowledge about the patient to increase robustness, accuracy, and reliability of the diagnosis. We use a Gibbs distribution to model the features we use for detection of each specific disc abnormality as well as domain knowledge incorporation. Gibbs distribution has been successfully used in statistical analysis and we find it flexible for incorporation of various pieces of information required for various diagnosis tasks. We study abnormality detection in general as a preprocessing step for focusing only on abnormal discs for further analysis. In this step, we only specify abnormal discs regardless of the type of abnormality. We then study two popular clinical disc degenerative diseases, i.e., disc desiccation and herniation. We incorporate appearance and context of discs in studying disc desiccation because it mainly related to relative signal levels of the discs. In herniation detection, we jointly incorporate shape and appearance information as herniation has major role in shape changes. We report our results with suitable validation metrics depending on the problem we are investigating.