Enabling quantitative lumbar degenerative disease auto-analysis with machine learning

Abstract

Lumbar degenerative disease (LDD) refers to a condition where the lumbar spine suffers from gradual wear and tear. It is more common in elderly populations. The LDD involves the deterioration of the lumbar vertebrae, cartilages, ligament, and especially intervertebral discs, in structure and function, which can lead to lower back pain, stiffness, limited mobility, numbness, and weakness. Current clinical LDD management highly relies on the manual assessment of specialists, which is usually subjective, laborious, and time-consuming, resulting in low clinical efficiency and consistency. In this thesis, a machine learning system for a comprehensive and quantitative auto-analysis of LDD is developed and validated, which enables multi-tissue segmentation, automated parameter measurement and pathology diagnosis, as well as LDD progression prediction. A dataset management strategy is first proposed for efficient archive and retrieval of the multi-dimensional clinical data for different analysis tasks. An unsupervised segmentation framework is further established for the multi-tissue segmentation of lumbar MRI. Based on the segmentation result, an economical model development process is developed to train the CNN model with limited training data. A CNN-based system is finally developed for comprehensive LDD assessment, including pathology diagnosis, progression prediction, and parameter measurement. A comprehensive validation is conducted on the system, the result demonstrates that: 1) our unsupervised segmentation framework can achieve accuracy comparable to the state-of-the-art fully supervised method without relying on any manual annotation; 2) the model development process enables efficient model training with a small training dataset; 3) the comprehensive LDD assessment system can achieve the accuracy close to clinical specialist in multiple tasks. The machine learning driven LDD assessment system can significantly improve the efficiency and consistency of clinical diagnosis for LDD and provide a meaningful reference for subsequent treatment planning, which will benefit both clinicians and patients.