How well does CardioViT-CAM generalize to echocardiographic data from different hospitals, ultrasound machines, and patient demographics, given that training and evaluation rely on a single-center private dataset?
That is an important concern. First, the model uses self-supervised Masked Autoencoder pretraining on 74,000 unlabeled echocardiography frames. This large-scale pretraining allows the network to learn fundamental cardiac structural patterns rather than memorizing dataset-specific characteristics. As a result, the learned features are more transferable to different clinical environments.
Second, the framework is based on a Vision Transformer architecture, which captures global anatomical relationships through attention mechanisms. Compared to traditional convolution-based models, transformers typically show improved adaptability to variations in imaging protocols, ultrasound devices, and scanning conditions.
Third, we applied extensive data augmentation, including geometric transformations, contrast variation, and noise simulation. These augmentations mimic real-world variability across hospitals, ultrasound machines, and patient populations, which improves model robustness.
In addition, Grad-CAM++ visualizations confirm that the model focuses on clinically meaningful cardiac defect regions rather than machine-specific artifacts, which further supports its reliability.
Although external multi-center validation is planned as future work, the combination of large-scale self-supervised learning, transformer-based feature extraction, and strong internal performance suggests that CardioViT-CAM has strong potential to generalize across diverse echocardiographic settings.
Currently, multi-center annotated CHD datasets are very limited. Therefore, we focused on large-scale self-supervised learning to compensate for data diversity. We are actively planning multi-institution collaborations for external validation.
