Fully connected network with multi-scale dilation convolution module in evaluating atrial septal defect based on MRI...

Area of Science:

• Medical Imaging
• Cardiovascular Diseases
• Artificial Intelligence

Background:

• Atrial septal defect (ASD) is a prevalent congenital heart disease resulting from abnormal embryonic development.
• ASD represents the largest proportion of congenital heart disease cases.
• Accurate segmentation of the atria is crucial for understanding and managing ASD.

Purpose of the Study:

• To design and implement an intelligent auxiliary segmentation system for ASD using deep learning.
• To develop a novel deep convolutional neural network for multi-scale feature extraction in cardiac MRI.
• To improve the accuracy of atrial segmentation for better ASD evaluation and surgical planning.

Main Methods:

• Proposed a multi-scale dilated convolution module with three parallel dilated convolutions.
• Incorporated a dense upsampling convolution module to mitigate information loss during upsampling.
• Retained the original FCN jump connection module for shallow position information and compared with K-means, Watershed, and U-net.

Main Results:

• The multi-scale dilated convolution network achieved superior segmentation results.
• The algorithm extracts associated pixel features across multiple ranges, capturing deeper information with limited downsampling.
• Achieved a Proportion of Greater Contour (PGC) of 98.78%, Average Perpendicular Distance (ADP) of 1.72mm, and Overlapping Dice Metric (ODM) of 0.935.

Conclusions:

• The multi-scale dilated convolution model significantly improves segmentation accuracy compared to other methods.
• This technique assists in ASD segmentation and defect extent evaluation.
• The system enhances surgical planning for atrial septal occlusion procedures.