Paraspinal Muscle Segmentation Based on Deep Neural Network.

Area of Science:

• Medical Imaging
• Computer Vision
• Spinal Surgery

Background:

• Accurate paraspinal muscle segmentation in Magnetic Resonance (MR) images is vital for automated analysis of lumbar diseases like chronic low back pain, disc herniation, and lumbar spinal stenosis.
• Existing automatic segmentation methods face challenges including unclear muscle boundaries, overlapping gray histogram distributions between target and background, and significant intra- and inter-patient shape variability.

Purpose of the Study:

• To develop an automated method for segmenting multifidus and erector spinae muscles in MR images.
• To address the limitations of current segmentation techniques by proposing a novel deep learning architecture.

Main Methods:

• A deformed U-net architecture was proposed, incorporating a residual module for gradient preservation and detail retention, and a feature pyramid attention (FPA) module for fusing multi-scale contextual information.
• The model was trained and validated using 120 labeled MR image cases provided by the Shengjing Hospital of China Medical University.

Main Results:

• The proposed model demonstrated high predictive capability in segmenting paraspinal muscles.
• Achieved a Dice coefficient of 0.949 and a Hausdorff distance of 4.62 mm for multifidus segmentation.
• Achieved a Dice coefficient of 0.913 and a Hausdorff distance of 7.89 mm for erector spinae segmentation.

Conclusions:

• The developed deformed U-net model effectively overcomes the challenges in paraspinal muscle segmentation.
• This advancement contributes to the development of automated measurement systems for paraspinal muscles, significantly aiding in the diagnosis and treatment of spinal diseases.