Towards Generalization of Medical Imaging AI Models: Sharpness-Aware Minimizers and Beyond

AI models have become ubiquitous tools of choice for different medical imaging problems like enhancement, work-flow acceleration, etc.. While availability of large amounts of diverse data and reliable annotations continue to be a challenge, development cycles of these models have shrunk. This necessitates a reliable recipe for improving generalization of AI models that fare well during deployment on unseen data. In this paper, we investigate generalization through the lens of sharpness-aware optimizers. We study two representative problems in medical imaging: (a) a difficult task of cardiac view classification on ultrasound images and (b) COVID-19 detection from chest X-ray images and demonstrate high efficacy of flat minima solutions. Further, we perform extensive Hessian analysis that reveals the impact of the geometry of loss landscape towards generalization. Our empirical studies suggest that sharpness aware minimization improves generalization by 5−10%, over and above the gain obtained by other methods - on both in-domain and out-of-domain test data.