| Summary: | <p>The objective of this thesis is the automation of radiological gradings in spinal lumbar Magnetic Resonance Images (MRIs). Solving this is extremely beneficial as this in a way would help in the standardization of gradings especially for back pain research. The output of the research done in this thesis would allow extremely fast readings of clinical scans which can potentially be useful in a large scale epidemiological study of spine-related diseases and aide clinical decision making.</p> <p>First, we build a pipeline to automatically produce radiological gradings of spinal lumbar MRIs and also localize the predicted pathologies. We show that: (i) a Con- volutional Neural Network (CNN) is able to predict multiple gradings at once, and we propose variants of the architecture including a multi-modal CNN that is able to take in both axial and sagittal or T1-weighted and T2-weighted scans; and (ii) a localization method that clearly shows pathological regions in the disc volumes using only a CNN trained for classification. The CNN is applied to a large corpus of standard clinical scan MRIs acquired from multiple machines via various scanning protocol, and is used to automatically compute intervertebral disc and vertebral body gradings for each MRI. We explore several radiological gradings: Pfirrmann grading, disc narrowing, upper/lower endplate defects, upper/lower marrow changes, spondy- lolisthesis, central canal stenosis, anterior/posterior disc bulging, and disc herniation. We report near human performances across all the gradings, and also visualize the evidence for these gradings localized on the original scans.</p> <p>Then, since a significant proportion of patients scanned in a clinical setting have follow-up scans; we show that such longitudinal scans alone can be used as a form of “free” self-supervision for training a deep network. We demonstrate this self- supervised learning for the case of T2-weighted sagittal lumbar MRIs. This learning via self-supervision can act as a pre-training regime when labelled data is sparse. We show that the performance of the pre-trained CNN on the supervised classification task is (i) superior to that of a network trained from scratch; and (ii) requires far fewer annotated training samples to reach an equivalent performance to that of the network trained from scratch.</p> <p>Finally, we show some preliminary results in mapping disc features learnt from radiological gradings to the Oswestry Disability Index (ODI) which is a measure of disability commonly used by back pain patients.</p>
|
|---|