ICML 2022

Timezone: »

Saliency Guided Adversarial Training for Tackling Generalization Gap with Applications to Medical Imaging Classification System

Xin Li · Yao Qiang · CHNEGYIN LI · Sijia Liu · Dongxiao Zhu

in New Frontiers in Adversarial Machine Learning »

This work tackles a central machine learning problem of performance degradation on out-of-distribution (OOD) test sets. The problem is particularly salient in medical imaging based diagnosis system that appears to be accurate but fails when tested in new hospitals/datasets. Recent studies indicate the system might learn shortcut and non-relevant features instead of generalizable features, so-called `good features'. We hypothesize that adversarial training can eliminate shortcut features whereas Saliency guided training can filter out non-relevant features; both are nuisance features accounting for the performance degradation on OOD test sets. With that, we formulate a novel model training scheme for the deep neural network to learn good features for classification and/or detection tasks ensuring a consistent generalization performance on OOD test sets. The experimental results qualitatively and quantitatively demonstrate the superior performance of our method using the benchmark CXR image data sets on classification tasks.

Author Information

Xin Li (Bosch AI)

Yao Qiang (Wayne State University)

CHNEGYIN LI (Wayne State University)

Sijia Liu (Michigan State University)

Dongxiao Zhu (Wayne State University)

Dongxiao Zhu is currently an Associate Professor at Department of Computer Science, Wayne State University. He received the B.S. from Shandong University (1996), the M.S. from Peking University (1999) and the Ph.D. from University of Michigan (2006). Dongxiao Zhu's recent research interests are in Machine Learning and Applications in health informatics, natural language processing, medical imaging and other data science domains. Dr. Zhu is the Director of Machine Learning and Predictive Analytics (MLPA) Lab and the Director of Computer Science Graduate Program at Wayne State University. He has published over 70 peer-reviewed publications and numerous book chapters and he served on several editorial boards of scientific journals. Dr. Zhu's research has been supported by NIH, NSF and private agencies and he has served on multiple NIH and NSF grant review panels. Dr. Zhu has advised numerous students at undergraduate, graduate and postdoctoral levels and his teaching interest lies in programming language, data structures and algorithms, machine learning and data science.