Automatic morphological classification of mitral valve diseases in echocardiographic images based on explainable deep...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Carpentier's functional classification categorizes mitral valve regurgitation based on morphology.
• The classification includes four types: Type I (normal), Type II (prolapse), Type IIIa (stenosis), and Type IIIb (restricted motion).
• Accurate classification is crucial for understanding and managing mitral valve pathologies.

Purpose of the Study:

• To develop an automated system for classifying mitral valve morphologies using echocardiographic images.
• To apply deep learning techniques for accurate and efficient mitral valve analysis.
• To provide explainability for the automated classification process.

Main Methods:

• Echocardiographic images (apical 4-chamber and parasternal long-axis views) were processed to extract cardiac cycle phases.
• Six pre-trained deep learning models were fine-tuned for mitral valve classification.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was used for visualization and explainability.

Main Results:

• The study achieved an 80% model accuracy on the test dataset.
• Specific architectures (Inception-ResNet-v2 and ResNeXt50) were identified as effective for classification tasks.
• Grad-CAM maps provided insights into the models' decision-making processes.

Conclusions:

• An explainable, automated, rule-based deep learning procedure for classifying mitral valve morphologies was developed.
• The findings suggest deep learning models can provide quick and precise assessments of mitral valve conditions.
• This study presents the first public dataset for Carpentier classification of mitral valve pathologies.