Automated Cardiovascular Pathology Assessment Using Semantic Segmentation and Ensemble Learning.

Area of Science:

• Cardiovascular Imaging
• Medical Image Analysis
• Machine Learning in Medicine

Background:

• Cardiac magnetic resonance imaging (CMR) offers high spatial resolution for cardiovascular disease staging.
• Manual segmentation and evaluation of cardiac cine sequences are time-consuming and prone to observer bias.
• Accurate quantification of cardiac function and morphology is crucial for disease assessment.

Purpose of the Study:

• To develop a fully automated processing pipeline for cardiac pathology assessment using CMR.
• To combine semantic segmentation and disease classification for efficient and reproducible analysis.
• To reduce the cost and observer variability associated with manual cardiac evaluations.

Main Methods:

• A robust dilated convolutional neural network was employed for voxel-wise segmentation of myocardium and ventricular cavities.
• A comprehensive volumetric feature matrix was generated from segmented data.
• A cardiac pathology classifier was modeled using the feature matrix for automated disease classification.

Main Results:

• The automated method achieved high segmentation accuracy with Dice index scores of 0.940 (left ventricle), 0.886 (myocardium), and 0.849 (right ventricle).
• A 5-ary pathology classification accuracy of 90% was achieved on an independent test set.
• The pipeline demonstrated potential for accurate, efficient, and reproducible cardiac pathological assessment.

Conclusions:

• The developed automated pipeline shows significant promise for improving cardiac pathological assessment.
• Machine learning techniques can enhance the efficiency and reproducibility of CMR analysis.
• This approach has the potential to streamline cardiovascular disease staging and management.