Algorithms for automatic localization, segmentation and diagnosis of lumbar pathology
MetadataShow full item record
Lower back pain is widely prevalent in people all over the world and negatively affects the quality of life due to chronic pain and change in posture. According to the American Academy of Orthopedic Surgeons (AAOS) four out of five adults experience low back pain at some point during their lives. The National Center for Health Statistics shows that more than 30 million MRI exams are conducted annually in the US and half of them are spine-related. A matter of great concern is that, in the last decade there has been a severe shortage of radiologists and projections show that there will be a significant boom in the ratio of their demand and supply. A CAD (Computer- Aided Diagnosis) system to generate diagnostic results from clinical lumbar (lower back) MR and CT scans would not only reduce the burden on a radiologist, but also boost the confidence on a diagnosis. This motivates us to strive towards the development of a robust, accurate and fully automated system to localize the various tissues like the vertebra, inter-vertebral discs and the spinal cord in clinical lumbar scans. A complete localization and segmentation would eventually be used by an automatic diagnostic system to detect lumbar abnormalities like disc herniation without any human intervention. Towards that end, my doctoral dissertation proposal consists of two parts : 1) Automatic Diagnosis of Lumbar Disorders : We propose algorithms for diagnosis of lumbar disorders using clinical CT and MRI scans. First, we present a fully automated method for robustly localizing and segmenting the lumbar vertebrae along with a method for diagnosis of wedge compression fracture from clinical lumbar CT. Second, we propose algorithms to circumvent the challenging issue of disc segmentation in MRI for disc herniation detection. Unlike past research, we present a localization method which combines supervised learning with heuristics to output a disc bounding box with enhanced localization accuracy. We also present feature extraction methods from the disc bounding boxes and report a performance comparison of individual and combined features for abnormal disc detection. Generally a radiologist detects an abnormality like disc herniation using the sagittal scans, and then makes use of the axial scans to localize and quantify, i.e., to estimate the location and size of the lumbar pathology. With this in mind, we finally present encouraging results on the use of axial MRI slices for automatic diagnosis, which could be extended for localization and quantification of disc abnormalities. 2) Complete segmentation of lumbar MRI : We propose algorithms for complete segmentation of sagittal lumbar MRI using two approaches--first, is an atlas-based approach, while the second is an auto-context based approach, both of which intelligently utilizes neighborhood label information along with appearance information for complete MRI segmentation. In the near future, we plan to utilize these algorithms towards the development of a completely automated and robust system to diagnose, locate and quantify lumbar disc disorders.