Automated quantification and phenotyping of lumbar intervertebral disc degeneration from clinical MRI with machine learning

Low back pain (LBP) affects a huge number of people worldwide, causing massive suffering. Due to the complexity of LBP, pinpointing the contribution of specific changes or damage is difficult. The discs between the vertebrae, which both cushion and support the spine, are one structure thought to play an important role in LBP as they commonly break down over time. Studying this requires the objective description of how the discs appear on routine clinical magnetic resonance imaging (MRI) scans. It is time-consuming work done by trained specialists, but machine learning has the potential to automate aspects of the process and improve the reliability of the measurements. Doing so would make it more practical to analyse far larger datasets. In this way we can better understand why some people develop LBP, how patients might be grouped by underlying cause, and what biological markers can signal risk. This thesis focuses on improving automated analysis of disc degeneration and extracting more useful information from routine lumbar spine MRI.

First, an existing artificial intelligence (AI) tool for classifying disc degeneration was tested against assessments made by trained radiologists. The study utilised the MRI scans of 1337 participants from the Northern Finland Birth Cohort 1966. The AI model was found to be reliable and applicable for large scale analyses, although there were some discrepancies compared to the radiologists. Next, the same MRI scans were analysed using an AI method to delineate the disc from the images (termed segmentation). An approach termed radiomics, applied to the segmentations, was used to measure the imaging characteristics of the discs. The resulting features were found to provide valuable new insights into the classification of disc degeneration from AI-generated disc segmentation applied to routine clinical MRI.

The final part of the thesis aimed to identify patterns in the accumulation of degenerative changes across the lumbar spine. The work utilised MRIs from a cross-sectional cohort of 423 patients with chronic LBP. Two distinct patterns were found, which also had a subtle relationship with the type of pain reported by the patients. The work showed that is is possible to model how spine degeneration unfolds over time in the absence of follow up data. This offers new opportunities to identify biological markers and underlying causes of LBP using large-scale datasets.