Denoising diffusion model for increased performance of detecting structural heart disease.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Cardiology

Background:

• Generative artificial intelligence (AI) shows potential for creating realistic medical images.
• Denoising diffusion probabilistic models are advanced generative AI techniques.
• Echocardiogram data combined with demographics presents complex distributions.

Purpose of the Study:

• To develop a generative AI model for synthetic echocardiogram data generation.
• To evaluate the diagnostic performance of deep learning models trained on synthetic data.
• To assess the impact of synthetic data augmentation on model accuracy, particularly for specific patient demographics.

Main Methods:

• A denoising diffusion probabilistic model was trained on the CheXchoNet dataset.
• The model encoded demographic data and echocardiogram measurements.
• A synthetic dataset was generated, enriched with younger patients and structural left ventricle disease cases.

Main Results:

• A deep learning model trained on synthetic data achieved comparable performance (AUROC=0.75) to one trained on real data.
• Combining real data with synthetic positive samples improved diagnostic accuracy (AUROC=0.80).
• The most significant performance gains were observed in younger patient subgroups.

Conclusions:

• Diffusion models can generate valuable synthetic medical data for AI training.
• Synthetic data augmentation enhances diagnostic deep learning model accuracy.
• Generative AI can help fine-tune models for specific patient populations, improving healthcare equity.