Expanding Domain-Specific Datasets with Stable Diffusion Generative Models for Simulating Myocardial Infarction.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Artificial intelligence (AI) accelerates human activity identification, but data scarcity hinders progress, especially in computer vision requiring extensive datasets.
• Training AI models for specialized or uncommon activities, such as fall detection or identifying heart attack symptoms, is challenging due to insufficient data.
• Existing datasets often lack the domain-specific images needed for accurate AI model training in critical applications.

Purpose of the Study:

• To propose a novel approach using generative models to augment image datasets for AI applications.
• To adapt stable diffusion models with low-rank adaptation for generating domain-relevant synthetic images.
• To address the challenge of data sparsity in AI-based computer vision tasks, particularly for identifying critical health events.

Main Methods:

• Developed a generative approach by refining stable diffusion models using low-rank adaptation.
• Created and annotated a dataset of 100 images depicting individuals simulating heart attack situations and neutral poses.
• Evaluated the generated synthetic images using learned perceptual image patch similarity (LPIPS) to assess their relevance to the target scenario.

Main Results:

• Demonstrated the potential of synthetically generated datasets to overcome data sparsity in AI applications.
• The proposed strategy effectively generated domain-relevant images for specialized computer vision tasks.
• Achieved promising results in creating a usable dataset for identifying critical health events through AI.

Conclusions:

• Synthetic datasets generated via the proposed method offer a cost-effective and ethically sound alternative to traditional data collection.
• This approach streamlines research by allowing researchers to own, modify, and expand datasets without additional permissions.
• The method shows significant potential for applications in smart environments, health monitoring, and anomaly detection, overcoming data limitations.