Medical imaging, in particular, has greatly benefited from recent advances in machine learning. Even with their exceptional performance, these models frequently don't provide all of the detail critical for medical providers to make informed decisions. In response, Google researchers created the innovative StylEx framework, which improves the explainability of AI models in medical imaging by utilizing generative AI. Conventional techniques, like heatmaps, highlight the significance of certain pixels, but fall short of elucidating more complex features like texture and form. By showing the visual signals connected with a classifier's predictions, StylEx, which is based on a StyleGAN2 image generator, aims to get beyond these restrictions and provide a more in-depth understanding of the "what" and "why" behind AI judgments.
The StylEx framework involves training a high-performing classifier on medical imaging data, ensuring relevant signal detection, and subsequently using a StyleGAN2-based generator to produce realistic images that preserve the classifier's decision-making process. An interdisciplinary panel, comprising clinicians, social scientists, and machine learning engineers, reviews these visualizations to interpret the attributes, identify potential biases, and generate hypotheses for future research. This collaborative approach not only enhances the explainability of AI models, but also ensures rigorous interpretation of findings, with the potential to uncover novel scientific insights and new avenues for inquiry in medical imaging.
