Local Label-Informed Feature Transfer for Generating Ground-Truth Medical Images: A Comparison of GAN- and Diffusion-Based Approaches

Synthetic medical images can help validate AI explanations without manual annotation burden: LLIFT generates realistic brain lesions in controlled locations without requiring pixel-level expert annotations during training, addressing a major bottleneck in creating ground-truth datasets for testing Explainable AI methods.

Two different AI approaches (GANs and diffusion models) both produced clinically realistic results: Both LLIFT-GAN and LLIFT-DM achieved quality scores comparable to the difference between real healthy and diseased images, suggesting multiple technical pathways can generate trustworthy synthetic medical data.

Semi-synthetic data could improve AI validation practices: By creating images with known, precisely-located lesions, researchers can better test whether XAI methods correctly identify which image features AI systems actually use for diagnosis—a critical step in ensuring medical AI systems are trustworthy.

This approach reduces reliance on expert labeling, which introduces human error: The framework avoids the accuracy problems inherent in having radiologists manually mark lesion locations, potentially making AI validation more scalable and consistent across institutions.

Realistic synthetic lesions matter more than artificial ones for testing AI systems: The paper demonstrates that hand-crafted artificial lesions lack clinical realism, so using AI-generated lesions that mimic actual pathology provides a more rigorous test of whether XAI methods work in real-world medical contexts.