Masked Autoencoders (MAE) are a commonly used self-supervised learning technique. However, in the context of your task, it remains unclear how MAE improves the performance of the downstream task. Could you please clarify this point through an ablation study or by comparing against a purely supervised learning pipeline?
Thank you for the comment. The role of MAE in our framework is to improve feature representation learning using large-scale unlabeled echocardiographic data. In the first stage, MAE pretraining enables the Vision Transformer to learn anatomical and spatial cardiac patterns by reconstructing masked image patches. These learned representations provide a stronger initialization for downstream supervised classification. Compared with purely supervised learning using only 6,000 labeled images, MAE leverages 74,000 unlabeled frames, which improves generalization and reduces overfitting. This advantage is reflected in our results, where the proposed MAE + ViT + Grad-CAM++ framework achieved 97.03% accuracy, outperforming previous fully supervised CNN-based approaches reported in the literature (86.5%–92.3%). Although a direct supervised-only ablation was not included in the current manuscript, the performance improvements and enhanced Grad-CAM++ localization demonstrate the effectiveness of MAE pretraining. We plan to include a detailed ablation comparison in future work.
