AI-driven MRI analysis for low back pain management

Download thesis

J. van der Graaf

  • Promotor: M. de Kleuver and B. van Ginneken
  • Copromotor: M. van Hooff and M. Rutten
  • Graduation year: 2025
  • Radboud University

Abstract

This thesis comprises several chapters on various topics related to automatic MRI analysis for patients with LBP. Given the large variety of pathologies that could potentially cause LBP, determining the most important aspects to analyze automatically is challenging. Each chapter addresses a specific facet of this complex issue.

Chapter 2 is a narrative review examining the relationship between LBP and specific pathologies that can be seen on MRI, known as image features. This review provides an overview of all relevant MR image features which have a proven relation to LBP. Accurate segmentation of the spine is essential for extracting specific image features.

In Chapter 3, we created a large multicenter segmentation dataset for vertebrae, intervertebral discs, and spinal canal segmentation. This dataset was made publicly available, except for a hidden test set. This test set was used to host a public continuous segmentation challenge to compare algorithms and set a benchmark. Lastly, this chapter presents two baseline segmentation algorithms for other researchers to use as a comparison.

In the subsequent chapters, we utilized the segmentation masks, automatically generated using the algorithm presented in Chapter 3, to classify or quantify features related to LBP.

Chapter 4 demonstrates our ability to quantify degenerative scoliosis by automatically measuring the Cobb angle across all vertebral levels. The performance of this method was compared to that of two radiologists and one spine surgeon.

Chapter 5 describes an algorithm that extracts specific measurements of the spinal canal at any given level, used to classify spinal canal stenosis using only sagittal images. This method was compared to expert radiologists who also had access to axial images of the spine.

In Chapter 6, we present a novel visualization method to evaluate degenerative disc disease. Along with the visualization, the method automatically measures disc height, disc bulging, and spondylolisthesis.

Chapter 7 summarizes the key contributions of this thesis, discussing implications and future directions.