Saliency Guided Adversarial Training for Learning Generalizable Features with Applications to Medical Imaging Classification System

Nevertheless, the performance degradation on OOD test sets remains a salient problem (Shao et al., 2020). One observation This work tackles a central machine learning is that the current approach introduces a nearly ideal problem of performance degradation on out-ofdistribution scenario for DNN to learn spurious shortcuts or non-relevant (OOD) test sets. The problem is particularly features (Geirhos et al., 2020) that do not exist in OOD test salient in medical imaging based diagnosis sets. In medical imaging systems, the problem becomes system that appears to be accurate but fails even more salient due to the significant distribution shift when tested in new hospitals/datasets. Recent between imaging data sets acquired from different hospitals, studies indicate the system might learn shortcut populations, and time periods. As a result, the AI imaging and non-relevant features instead of generalizable system that is seemingly effective on training sets often does features, so-called'good features'. We hypothesize not generalize well to new hospitals or data sets (DeGrave that adversarial training can eliminate shortcut et al., 2021). Fortunately, in the relatively closed medical features whereas saliency guided training can imaging environment, we are not so much concerned about filter out non-relevant features; both are nuisance adversarial OOD test sets. Instead, we consider how to features accounting for the performance degradation leverage adversarial IID data sets for learning good features.