3D multi-scale feature extraction and recalibration network for spinal structure and lesion segmentation.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Spinal diagnostics

Background:

• Automatic segmentation is a key technique for diagnosing spinal conditions.
• Accurate segmentation of spinal structures is crucial for effective diagnosis and treatment planning.

Purpose of the Study:

• To develop and assess a deep convolution network for segmenting key spinal components on MRI scans.
• The network aims to segment the intervertebral disc, spinal canal, facet joint, and herniated disc.

Main Methods:

• A novel deep neural network incorporating 3D squeeze-and-excitation and multi-scale feature extraction blocks was designed.
• Weighted cross-entropy loss was used to handle class imbalance during training.
• Semi-supervised segmentation was employed to minimize the need for manual annotation.

Main Results:

• The proposed model achieved a 77.67% mean intersection over union (IoU).
• It demonstrated significant performance gains (9.56% and 11.11%) over V-Net and U-Net architectures.
• The semi-supervised approach proved effective in reducing annotation labor.

Conclusions:

• The developed 3D multi-scale feature extraction and recalibration network excels in segmenting spinal structures and herniated discs.
• This advanced network outperforms traditional encoder-decoder networks in segmentation accuracy.
• The study highlights the potential of deep learning for improved spinal condition diagnosis.