Automatic quantification of the LV function and mass: A deep learning approach for cardiovascular MRI

Area of Science:

• Cardiovascular Imaging Analysis
• Artificial Intelligence in Medicine
• Deep Learning for Medical Image Segmentation

Background:

• Accurate quantification of the left ventricle (LV) is crucial for diagnosing and managing cardiovascular diseases.
• Manual segmentation and quantification of the LV from medical images are time-consuming and prone to inter-observer variability.
• Deep learning, particularly convolutional neural networks (CNNs), offers a promising avenue for automating these complex tasks.

Purpose of the Study:

• To propose a novel deep learning framework for automatic left ventricle (LV) quantification.
• To develop and evaluate new CNN architectures for enhanced LV segmentation and functional parameter estimation.
• To compare the performance of the proposed method against manual segmentation variability.

Main Methods:

• A framework utilizing two CNNs: one for LV detection and another for tissue classification.
• Introduction of three novel deep learning architectures incorporating sparsity, depthwise separable convolutions, and residual learning within a U-net framework.
• Employing the generalized Jaccard distance as the optimization objective function for training.

Main Results:

• The proposed CNNs achieved high accuracy in myocardial segmentation, with a Dice's coefficient of approximately 0.9.
• Strong correlations were observed with key physiological measures: 0.99 for end-diastolic/end-systolic volume, 0.97 for myocardial mass, and 0.95 for ejection fraction.
• The method demonstrated excellent performance for stroke volume (0.93) and cardiac output (0.93) quantification.

Conclusions:

• The developed CNN-based approach effectively quantifies structural and functional LV features, including LV mass and ejection fraction (EF).
• The proposed method shows capability and merits for aiding in the diagnosis and treatment of various cardiovascular pathologies.
• The automated LV quantification errors are comparable to inter- and intra-operator ranges for manual contouring, indicating clinical feasibility.