Reproduce “Feature Extraction for Generative Medical Imaging Evaluation: New Evidence Against an Evolving Trend” on OpenResearch

About this paper

McKell Woodland, Austin Castelo, Mais Al Taie, Jessica Albuquerque Marques Silva, Mohamed Eltaher, Frank Mohn, Alexander Shieh, Suprateek Kundu and 3 more

This study challenges the assumption that medical image-trained feature extractors improve Fréchet Distance reliability for evaluating synthetic medical images, finding that ImageNet-based extractors, particularly self-supervised ones, align better with human judgment. The research also demonstrates that Differentiable Augmentation consistently produces hyper-realistic synthetic medical images.

Key insights

• ImageNet-based feature extractors, especially self-supervised ones like SwAV, produce more consistent generative model rankings that align better with expert human judgment in medical imaging than RadImageNet-based extractors.
• Domain-specific pre-training for FD feature extractors (e.g., RadImageNet) does not inherently improve metric validity for medical imaging and can lead to rankings inconsistent with human perception.
• Differentiable Augmentation (DiffAugment) is a highly effective data augmentation technique for generative medical imaging, consistently producing hyper-realistic images that can be perceived as more realistic than actual real images.

Summary by alphaXiv.