Comparison of Deep Learning Architectures for Cardiac Contour Segmentation in Catheterization Radiographs

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Accurate cardiac image segmentation is crucial for assessing heart anatomy and function.
• Manual segmentation is time-consuming and prone to variability.
• Deep learning models like U-Net and DeepLabV3 show potential for automating cardiac segmentation.

Purpose of the Study:

• To compare the performance of U-Net and DeepLabV3 for cardiac silhouette segmentation on cardiac catheterization radiographs.
• To evaluate the effectiveness of automated segmentation in improving efficiency and reproducibility of cardiac assessments.

Main Methods:

• A supervised deep learning approach using 1717 chest radiographs from cardiac catheterization.
• Implementation and comparison of a modified U-Net and a DeepLabV3 network.
• Evaluation using Dice similarity coefficient, Intersection over Union (IoU), and pixel accuracy.

Main Results:

• U-Net achieved higher mean Dice (0.9454) and IoU (0.8980) compared to DeepLabV3 (Dice=0.9321, IoU=0.8742).
• Statistical analysis confirmed U-Net's significantly superior performance across all metrics with large effect sizes.
• Pixel accuracy was also higher for U-Net (0.9844) versus DeepLabV3 (0.9806).

Conclusions:

• U-Net demonstrates superior accuracy for cardiac silhouette segmentation on catheterization radiographs.
• The encoder-decoder architecture with skip connections in U-Net facilitates precise boundary delineation.
• Automated segmentation using U-Net can enhance cardiomegaly detection and cardiac monitoring efficiency.