Automated radiological grading of spinal MRIs via weak and self-supervision

Thesis

The main aim of this thesis is to create automated tools that can interpret spinal MRI scans as accurately as an expert radiologist. This is motivated by the need to reduce waiting times for scan interpretation, address the issue of high variability among readers for common grading methods, and facilitate the scaling of clinical studies to larger cohorts. We begin by exploring automated vertebrae detection in clinical scans. First, we propose a sequential detection method, where the location of previous vertebra detection is used as a prior for the next. This is followed by an improved, language-modelling-inspired method that allows for rapid, accurate detection and labelling in arbitrary sequence and field-of-view. Next, we propose a novel transformer-based architecture for diagnosis of spinal disorders, demonstrating the importance of incorporating contextual information from surrounding vertebrae and multiple MR sequences in grading decisions. This is tested on disc degeneration and cancer datasets, yielding strong results in both cases.

Related prior research has relied on manually annotated data almost exclusively for model training. However, this has several limitations (e.g. significant time commitments and domain knowledge required from annotators, high inter-reader variability). Therefore, we set an auxiliary goal; to reduce the need for manual annotation in medical imaging through self-supervised learning. To this end, we propose a novel method for learning representations of same-subject, differentmodality images pairs, using registration as a pretext task. This method is used to perform 3D spinal segmentation in MRIs using 2D DXA-scan annotations for supervision. Finally, we explore using free-text radiology reports as a source of weak supervision for representation learning. Experimenting with a dataset of chest Xrays, we adapt vision-language model training methods to improve performance when training data is limited, as is often the case in medical imaging.

Authors: Windsor, RL

• DOI: 10.5287/ora-vgxnk6zv9
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: medical imaging; self-supervised learning; spinal MRI; deep learning