Automatic echocardiographic anomalies interpretation using a stacked residual-dense network model.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Echocardiographic interpretation is crucial for diagnosing cardiac septal abnormalities but is time-consuming and prone to human error.
• Current automatic segmentation methods identify abnormality regions but lack defect position classification for complete interpretation.

Purpose of the Study:

• To develop a stacked residual-dense network model for automatic segmentation and defect position classification of cardiac abnormalities in echocardiograms.
• To create a generalized model applicable to both prenatal and postnatal echocardiography.

Main Methods:

• A stacked residual-dense network was employed for segmentation and classification.
• The model was trained on 1345 echocardiograms and validated on 181 unseen echocardiograms.
• Performance was evaluated using Intersection over Union (IoU), Dice Similarity Coefficient (DSC), mean Average Precision (mAP), accuracy, specificity, and sensitivity, with validation by five cardiologists.

Main Results:

• The model achieved 58.17% IoU, 75.75% DSC, and 76.36% mAP on validation data.
• On unseen data, performance metrics were 42.39% IoU, 55.72% DSC, and 51.04% mAP.
• Defect position classification on unseen data showed high accuracy (92.27%), specificity (94.33%), and sensitivity (92.05%).

Conclusions:

• The proposed AI model demonstrates potential as a supporting diagnostic tool in clinical practice for echocardiographic interpretation.
• The model's ability to segment cardiac abnormalities and classify defect positions enhances diagnostic capabilities.
• Validation against human experts indicates promising suitability for clinical integration.