Curriculum learning from patch to entire image for screening pulmonary abnormal patterns in chest-PA X-ray: intra- and extra-validations on multi-center datasets

Beomhee Park, Joon Beom Seo, Sang Min Lee, Kyunghee Lee, Eun Sol Lee, Young Hoon Cho, Yongwon Cho, Namkug Kim

Abstract: In applying the deep-learning method to medical images, there is always the lack of data, and high dimensionality and complexity of medical images make this problem even more serious. Although chest X-ray image is two-dimensional data, accurately detecting abnormal patterns is a very difficult task due to its intrinsic limitations. Therefore, we proposed a computer-aided detection (CAD) for detecting 5 kinds of pulmonary abnormalities in chest-posterioranterior (PA) X-ray images with a curriculum learning strategy to train complex problem after training relatively easy problem to guide the CAD toward better local minima. In addition, extra-validation using multi-center datasets was performed to demonstrate the accuracy and robustness of this strategy.

Keywords: CAD, chest X-ray, multi-center, deep-learning, curriculum learning

Author Affiliation: University of Ulsan College of Medicine, Asan Medical Center, Seoul National University Bundang Hospital, South Korea