Generative Artificial Intelligence Enhancements for Reducing Image-based Training Data Requirements.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Data Science

Background:

• Intensive data requirements for AI models are a significant bottleneck, particularly in healthcare where data scarcity can lead to bias.
• Under-privileged social groups may face disadvantages due to limited access to healthcare data for AI training.

Purpose of the Study:

• To develop an autoencoder-based framework to augment scarce datasets and reduce reliance on large data volumes.
• To enhance the performance of AI models in healthcare by addressing data limitations.

Main Methods:

• A computational study utilizing open-source datasets from Singapore, China, India, and Spain (patients aged 40-80).
• An autoencoder was employed to generate synthetic optic disc images from real-world patient data.
• The expanded datasets were used to train AI models for detecting glaucomatous optic neuropathy.

Main Results:

• Enhancing training datasets with autoencoder-generated synthetic images resulted in superior datasets.
• The improved datasets led to enhanced performance of AI models in detecting glaucomatous optic neuropathy.

Conclusions:

• The developed framework effectively addresses the challenges of data volume and quality in AI model development.
• Findings have implications for advancing AI adoption in data-challenged fields beyond healthcare.