Automatic assessment of spinal deformities in 2D and 3D

Thesis

Scoliosis assessment still relies on 2D radiographs and manual measurements by expert clinicians. This thesis proposes deep learning pipelines that automate 2D measurements on dual energy X-ray absorptiometry (DXA) and 3D measurements on Magnetic Resonance Imaging (MRI), unifying 2D and 3D information.

We start by the development of an end-to-end pipeline for 2D scoliosis assessment using DXA scans from the UK Biobank. This foundational work introduces a U-Net style network that segments six body parts including the spine, coupled with an iterative label refinement loop that effectively suppresses segmentation failures. Through cubic spline fitting, we derive precise spine curvature metrics, establishing the spinal geometric framework that underpins the following work in this thesis.

We then shift to a 3D perspective by leveraging volumetric spine MRIs from the UK Biobank. Here, we develop a custom 3D U-Net-like model capable of accurately extracting spine meshes given limited annotation of the spine. This advancement enables comprehensive 3D shape analysis of the spine, capturing critical elements invisible in two dimensions, such as axial rotation and lordosis. The result is a richer set of biomarkers that enhance classification of spine deformity and scoliosis severity.

Building on these advancements, we bridge the gap between imaging modalities by pioneering a technique to predict 3D spine shape directly from a single DXA scan. Our approach employs a vision transformer with a regression head that outputs spine curves from a cropped DXA image, allowing us to reconstruct the full 3D vertebral stack with sub-millimeter accuracy. This image-to-shape model creates opportunities for 3D insights in clinical settings using only low-dose DXA systems.

We also develop an Automated DXA Scoliosis Method (DSM) that directly outputs the maximum angle of the spine, enabling fair comparison with human annotation. This automated approach is validated on manually annotated DXA scans by expert clinicians with strong agreement against expert annotations. The final stage of our research focuses on enhancing 3D segmentation of the spine to achieve vertebrae-level prediction. We address the challenge of generalizing segmentation to out-of-domain datasets by implementing test-time-adaptation techniques, further improving the robustness and clinical applicability of our methods.

Authors: Bourigault, E

• DOI: 10.5287/ora-zbb7rybyn
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: DXA; scoliosis; MRI; spine geometry; 2D-3D symbiosis
• Subjects: Engineering; Imaging; Imaging systems in medicine; Computer programming; Three-dimensional imaging