Abstract
Training robust AI models for medical diagnosis requires massive amounts of high-quality, labeled data, which is particularly scarce in cardiology due to clinical rarity and strict privacy regulations. This paper presents Echo-Path, a novel generative framework that utilizes diffusion models to synthesize realistic echocardiogram videos conditioned on specific heart pathologies, such as Atrial Septal Defect (ASD) and Pulmonary Arterial Hypertension (PAH). By augmenting real-world datasets with these synthetic videos, we demonstrate a significant improvement in diagnostic accuracy—up to 8%—surpassing existing state-of-the-art specialized architectures trained only on real data.
Introduction
In the domain of medical imaging, the scarcity of labeled data for rare or complex conditions is a major bottleneck for AI development. Echocardiograms for conditions like Atrial Septal Defect (ASD) or Pulmonary Arterial Hypertension (PAH) are difficult to acquire. While previous generative approaches like GANs have been explored, they often suffer from poor visual fidelity and temporal inconsistency. Echo-Path addresses these challenges by providing fine-grained control over the generation process, allowing for the creation of videos that accurately reflect the structural and motion markers of targeted diseases.
Methodology
Echo-Path employs a multi-stage, class-conditioned diffusion pipeline designed to generate high-fidelity, temporally consistent cardiac videos.
- Latent Image Diffusion Model (LIDM): The model operates in a compressed latent space via a Variational Autoencoder to generate a representative "seed" frame conditioned on a disease label (e.g., "ASD").
- Privacy Filtering: A re-identification (ReID) model ensures patient privacy by discarding generated frames that are too similar to real training data, preventing the model from simply memorizing patient records.
- Latent Video Diffusion Model (LVDM): This component extends the seed frame into a 64-frame sequence, ensuring the heart's motion reflects the specific pathology, such as the interventricular septal flattening characteristic of PAH.
- Autoregressive Sampling: To generate long-form videos spanning multiple heartbeats, a blockwise strategy is used, where the final frame of one segment serves as the starting point for the next, ensuring smooth temporal transitions.
Datasets and Experimental Setup
The framework was evaluated using two primary datasets: Cardiac-ASD (207 apical-4-chamber videos) and Cardiac-PAH (448 videos). All videos were standardized to a 112x112 resolution at 32 frames per second to ensure consistency across the training and evaluation phases.
Results and Discussion
Visual Fidelity and Pathology Representation
Echo-Path achieves superior visual fidelity compared to previous state-of-the-art models like EchoNet-Synthetic. It achieved a lower Fréchet Inception Distance (FID) score of 18.35 (compared to 20.05), indicating that the synthetic frames are statistically closer to real ultrasound data. Crucially, the model captures key clinical markers that are often missed by simpler generative models.
Diagnostic Performance
One of the most significant findings is the "Synthetic-to-Real" bridge. A classifier trained entirely on synthetic data achieved over 80% accuracy on real-world patients, suggesting the model captures underlying biological features rather than visual noise. Furthermore, augmenting real training sets with Echo-Path synthetic videos improved diagnostic performance for ASD by 7% and PAH by 8%. The augmented ResNet model reached a state-of-the-art accuracy of 91.8% for ASD detection, outperforming specialized architectures like CardiacNet trained only on real data.
Conclusion
Echo-Path demonstrates the potential of pathology-conditioned diffusion models to overcome data scarcity in medical imaging. While structural defects like ASD were modeled with high precision, subtle motion-based conditions like PAH remain more challenging due to the subtle variations found in real human hearts. Future work will focus on expanding the framework to include more diverse cardiac views and a wider range of pathologies, moving toward a universal generative model for echocardiography.
Summary updated