Explainable AI for Generative Biomedical Imaging: Visualizing the Synthetic-to-Real Gap

Fu et al. (2025, XAI-PSSGAN) show the promise of GANs for synthesizing clinically useful images in domains where real data is scarce or high-risk (e.g., NIR-IIb imaging). However, there is little work on explainability for these generative models: How do clinicians know what aspects of a synthetic image are trustworthy, and which features might be artifacts? This research would develop novel XAI techniques (potentially leveraging causal or counterfactual frameworks) to highlight differences between synthetic and real images, and to explain how these differences might impact clinical interpretation or subsequent AI model decisions. This is especially important as generative models become more common in clinical pipelines. The project would bridge the gap between generative model evaluation and clinical trust, providing much-needed transparency in an emerging area.

References:

1. XAI-PSSGAN: Perception-Enhanced Spectrum Shift Generative Adversarial Network with Explainable AI System for NIR-II Fluorescence Molecular Imaging. Lidan Fu, Binchun Lu, Lingbing Li, Xiaojing Shi, Jie Tian, Zhenhua Hu (2025). IEEE International Symposium on Biomedical Imaging.
2. XAI-PSSGAN: Perception-Enhanced Spectrum Shift Generative Adversarial Network with Explainable AI System for NIR-II Fluorescence Molecular Imaging. Lidan Fu, Binchun Lu, Lingbing Li, Xiaojing Shi, Jie Tian, Zhenhua Hu (2025). IEEE International Symposium on Biomedical Imaging.
3. XAI-PSSGAN: Perception-Enhanced Spectrum Shift Generative Adversarial Network with Explainable AI System for NIR-II Fluorescence Molecular Imaging. Lidan Fu, Binchun Lu, Lingbing Li, Xiaojing Shi, Jie Tian, Zhenhua Hu (2025). IEEE International Symposium on Biomedical Imaging.