Enhanced dataset synthesis using conditional generative adversarial networks.

Area of Science:

• Biomedical data science
• Machine learning in healthcare
• Medical imaging analysis

Background:

• Acquiring sufficient biomedical data for training is challenging and resource-intensive.
• Computer-aided diagnosis (CAD) algorithm performance is highly dependent on sample size and feature space.
• Limited datasets can hinder the development and accuracy of diagnostic tools.

Purpose of the Study:

• To propose a Conditional Generative Adversarial Network (CGAN) for enhanced feature generation and synthetic dataset creation.
• To address the difficulties in biomedical data acquisition and improve the success rates of CAD algorithms.
• To generate synthetic datasets with high class separability, enabling larger sample sizes for model training.

Main Methods:

• Utilized a Conditional Generative Adversarial Network (CGAN) to synthesize enhanced feature vectors and create large sample datasets.
• Trained the CGAN model using 25% of five diverse medical datasets.
• Evaluated and compared the performance of synthetic datasets against original data across various sample sizes.

Main Results:

• The proposed CGAN model successfully generated synthetic datasets with enhanced class separability.
• Generated datasets demonstrated the potential to increase the accuracy rates of CAD systems.
• Physicians can benefit from reduced sample collection efforts through CGAN-based data synthesis.

Conclusions:

• Conditional Generative Adversarial Networks offer a viable solution for augmenting limited biomedical datasets.
• The CGAN approach enhances feature vectors, leading to improved CAD system performance.
• This method significantly reduces the time and effort required for participant recruitment and data collection in medical research.