Deep learning-based structure segmentation and intramuscular fat annotation on lumbar magnetic resonance imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Spinal Diagnostics

Background:

• Lumbar disc herniation (LDH) is a common cause of low back pain.
• LDH patients often exhibit paraspinal muscle atrophy and fatty infiltration (FI), worsening symptoms.
• Magnetic Resonance Imaging (MRI) is vital for assessing paraspinal muscle condition.

Purpose of the Study:

• To develop a dual-model for automated muscle segmentation and FI annotation on lumbar spine MRI.
• To assist clinicians in comprehensively evaluating LDH conditions and associated muscle changes.

Main Methods:

• Retrospective collection of LDH patient data (December 2020 - May 2022).
• Development and validation of a dual-model for muscle segmentation and FI annotation.
• Performance evaluation using metrics like Dice Similarity Coefficient (DSC), Average Precision (AP), recall, F1 score, Cohen's Kappa, and Mean Absolute Percentage Error (MAPE).
• Comparison of model-derived FI measurements against threshold algorithms using MAPE.

Main Results:

• The muscle segmentation model achieved a DSC of 0.92 and AP of 0.98 on the internal test set.
• The fat annotation model attained a recall of 91.30% and F1 Score of 0.82.
• The developed model demonstrated lower MAPE for FI measurements compared to threshold algorithms across paraspinal muscles.

Conclusions:

• The developed dual-model shows excellent performance in automated muscle segmentation and FI annotation on lumbar spine MRIs.
• The model offers a more accurate method for quantifying fatty infiltration in paraspinal muscles compared to existing thresholding techniques.
• This tool can aid clinicians in better assessing LDH-related muscle changes and improving patient management.