Classification of lumbar spine disorders using large language models and MRI segmentation

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Spinal Diagnostics

Background:

• Magnetic Resonance Imaging (MRI) is crucial for diagnosing lumbar spine disorders, but its complexity can impede diagnostic accuracy.
• Current diagnostic methods face challenges due to the intricate nature of lumbar spine conditions.
• This study addresses the need for improved precision in classifying lumbar spine disorders.

Purpose of the Study:

• To develop and evaluate a BERT-based large language model (LLM) for enhanced classification of lumbar spine disorders.
• To integrate multimodal data, including MRI scans, textual reports, and numerical measurements, for improved diagnostic accuracy.
• To leverage advanced AI techniques for precise anatomical feature extraction and disorder classification.

Main Methods:

• MRI data segmentation quality was assessed using Dice coefficients and IoU metrics.
• A Convolutional Neural Network (CNN) extracted key features like lumbar lordotic angle and disc heights.
• A BERT-based spinal LLM integrated CNN-extracted MRI features and numerical values via early fusion, trained on 28,065 patient cases.

Main Results:

• The BERT-based LLM demonstrated high performance, with key metrics approaching 0.9 in classifying various lumbar spine disorders.
• External validation on 514 expert-validated cases confirmed the model's clinical relevance and generalizability.
• The model effectively classifies 61 distinct combinations of lumbar spine disorders, including spondylolisthesis, herniated disc, and spinal stenosis.

Conclusions:

• The BERT-based spinal LLM significantly enhances the precision of lumbar spine disorder classification.
• This AI-driven approach supports more accurate diagnoses and improved treatment planning for spinal conditions.
• The study highlights the potential of integrating LLMs with medical imaging for advanced diagnostic capabilities.