Recognition of Normal Fetal Echocardiograms Based on an Explainable Denoising Deep Learning Model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Fetal echocardiogram interpretation is crucial for diagnosing congenital heart defects.
• Deep learning models offer potential for automated analysis but often lack interpretability.
• Existing models may struggle with image quality variations in ultrasound data.

Purpose of the Study:

• To evaluate the Grouped Shared Convolutional Attention Vision Transformer (GSCAViT), an explainable denoising deep learning model.
• To assess GSCAViT's performance in classifying normal fetal echocardiograms.
• To improve interpretability and image quality in fetal cardiac ultrasound analysis.

Main Methods:

• A retrospective study analyzed 2501 images from 358 fetal cardiac ultrasound exams.
• GSCAViT, incorporating a denoising-guided GSCA module, was compared against baseline and enhanced models.
• Performance was evaluated using accuracy, precision, recall, F1 score, and SHAP for feature visualization.

Main Results:

• GSCAViT achieved high accuracy (97.1% on validation, up to 99.4% on test sets) with low error rates.
• SHAP visualizations identified critical cardiac structures, enhancing model interpretability.
• The denoising module improved image quality, evidenced by superior contrast-to-noise and peak signal-to-noise ratios.

Conclusions:

• GSCAViT demonstrated superior performance in classifying normal fetal echocardiograms compared to other models.
• Explainable AI through SHAP visualization improved the interpretability of the classification process.
• The denoising-guided GSCA module proved effective in enhancing image quality and classification efficacy.