Uncovering Bias in Medical Data Synthesis Through Visualization Technique

Acquiring and utilizing medical imaging data presents challenges such as difficulties in data collection and privacy protection. While generative models offer promising solutions for medical data synthesis, their effectiveness in clin-ical settings is often limited by potential generation bias. This necessitates a systematic analysis to ensure reliable applications across diverse clinical scenarios. To tackle this issue, we employ a visual analysis method to convey generation bias in medical images through distribution comparison. We first employ an intuitive visualization to display bias in generative models during medical data synthesis. Then, we conduct a detailed analysis of the distributional difference between generated and original data to extract posterior knowledge of bias. Our experimental results on two medical datasets demonstrate that this analysis method facilitates bias pattern comparison across models and helps identify potential factors that impact the quality of generated data.