A Machine Learning and Computer Assisted Methodology for Diagnosing Chronic Lower Back Pain on Lumbar Spine Magnetic Resonance Images

Chronic Lower Back Pain (CLBP) is one of the major types of pain that affects many people around the world. It is estimated that 28.1% of US adults suffer from this illness and 2.5 million of the UK population experience this type of pain every day. Most CLBP cases do not happen overnight and it is usually developed from a less serious but acute variant of lower back pain. An acute type of lower back pain can develop into a chronic one if the underlying cause is serious and left untreated. The longer a person is disabled by back pain, the less chance he or she returns to work and the more health care cost he or she will require. It is therefore important to identify the cause of back pains as early as possible in order to improve the chance of patient rehabilitation. The speediness of early diagnosis can depend on many factors including referral time from a general practitioner to the hospital, waiting time for a specialist appointment, time for a Magnetic Resonance Imaging (MRI) scan and time for the analysis result to come out. Currently diagnosing the lower back pain is done by visual observation and analysis of the lumbar spine MRI images by radiologists and clinicians and this process could take up much of their time and effort. This, therefore, rationalizes the need for a new method to increase the efficiency and effectiveness of the imaging diagnostic process. This thesis details a novel methodology to automatically aid clinicians in performing diagnosis of CLBP on lumbar spine MRI images. The methodology is based on the current accepted medical practice of manual inspection of the MRI scans of the patient’s lumbar spine as advised by several practitioners in this field. The main methodology is divided into three sub-methods the first sub-method is disc herniation detection using disc segmentation and centroid distance function. While the second sub-method is lumbar spinal stenosis detection via segmentation of area between anterior and posterior (AAP) Elements. Whereas, the last sub-method is the use of deep learning to perform semantic segmentation to identify regions in the MRI images that are relevant to the diagnosis process. The method then performs boundary delineation between these regions, identifies key points along the boundaries and measures distances between these points that can be used as an indication to the health of the lumbar spine. Due to a limitation in the size and suitability of the currently existing open-access lumbar spine dataset necessary to train and test any good classification algorithms, a dataset consisting of 48,345 MRI slices from a complete clinical lumbar MRI study of 515 symptomatic back pain patients from several specialty hospitals around the world has been created. Each MRI study is annotated by expert radiologists with notes regarding the observed characteristics, condition of the lumbar spine, or presence of diseases. The ground-truth dataset containing manually labelled segmented images has also been developed. To complement this ground-truth dataset, a novel method of constructing and evaluating the suitability of ground truth data for lumbar spine MRI image segmentation has been developed. A subset of the dataset, which includes the data for 101 patients, is used in a set of experiments that have been conducted using a variety of algorithms to conclude with using SegNet as the image segmentation algorithm. The network consists of VGG16 layers pre-trained using a subset of non-medical images from the ImageNet database and fine-tuned using the training portion of the ground-truth dataset. The results of these experiments show the accurate delineation of important boundaries of regions in lumbar spine MRI. The experiments also show very close agreement between the expert radiologists’ notes on the condition of a lumbar spine and the conclusion of the system about the lumbar spine in the majority of cases.