ABSTRACT Purpose To evaluate the proposed explainable denoising deep learning model, Grouped Shared Convolutional Attention Vision Transformer (GSCAViT), for classifying normal fetal echocardiogram. Methods A retrospective study was conducted on 358 fetal cardiac ultrasound exams with a total of 2501 images. GSCAViT utilized seven echocardiograms and was compared against baseline and enhanced models using metrics including accuracy, precision, recall, F1 score. The SHAP method clarified key image features, and the denoising guided GSCA module was assessed through visual comparisons and image quality metrics. Results GSCAViT achieved an accuracy of 97.1%, 99.4%, 81.3%, 72.9% on the validation and three test sets. In addition, GSCAViT achieved low error rates of 2.9%, 0.6%, 18.7%, and 27.1%. To improve upon existing Vision Transformer based models and denoising modules, we propose the GSCAViT, which integrates a novel denoising‐guided GSCA module for enhanced image quality and interpretability. SHAP visualizations confirmed the model's ability to identify critical cardiac structures, while the denoising module enhanced image quality, yielding the highest contrast‐to‐noise ratio and peak signal‐to‐noise ratio values. Conclusion GSCAViT outperformed baseline and several enhanced models in classifying seven types of normal fetal echocardiograms, SHAP visualization enhanced the interpretability of classification, comparisons of visual effectiveness and image parameters confirmed the efficacy of the GSCA module.
Song et al. (2026) studied this question.