Cardiac Magnetic Resonance Image Segmentation Method Based on Multi-Scale Feature Fusion and Sequence Relationship...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Accurate left atrial segmentation is crucial for diagnosing atrial fibrillation (AF) and planning robotic surgery.
• Cardiac MRI presents challenges due to varying anatomical scales and slice-to-slice continuity.

Purpose of the Study:

• To develop an advanced image segmentation method for left atrial structures in cardiac MRI.
• To enhance diagnostic capabilities for AF and support robotic surgical interventions.

Main Methods:

• A novel deep learning approach combining sequence relationship learning and multi-scale feature fusion.
• Utilized convolutional neural network layers with attention modules for feature extraction and fusion.
• Employed recurrent neural network layers to capture inter-slice correlations in cardiac MRI sequences.

Main Results:

• Achieved Dice scores of 90.73% and 92.05% on LASC2013 and ASC2018 datasets, respectively.
• Reported IoU values of 89.37% and 89.41% on the LASC2013 and ASC2018 datasets.
• Demonstrated Hausdorff distances of 4.803 mm and 9.056 mm on the LASC2013 and ASC2018 datasets.

Conclusions:

• The proposed method effectively segments left atrial structures in cardiac MRI.
• This technique offers improved accuracy for AF diagnosis and robotic surgery planning.
• The approach addresses challenges of multi-scale features and sequential data in cardiac imaging.