Published
Union Conference on Lung Health (2022)
Authors
Heejun Shin1, Juhyung Park1, Taehee Kim1, Jongho Lee2, Hyungjin Kim3, and Dongmyung Shin1
Affiliations
1AI Engineering Division, Radisen Co. Ltd., Seoul, KOREA, REPUBLIC OF,
2Seoul National University, Seoul, KOREA, REPUBLIC OF,
3Seoul National University Hospital,
Seoul, KOREA, REPUBLIC OF.
Background
Recently, many deep learning-based tuberculosis (TB) screening methods for chest X-rays (CXRs) have been proposed. However, network trained using CXRs from one domain provides reduced performances when applied to CXRs of different domains (e.g., vendors and institutions). Here, we propose domain agnostic image processing (DAIP) pipeline based on knowledge of X-ray physics to minimize performance degradation.
Design/Methods
[Datasets] CXRs from Vietnam national lung hospital (digital radiography (DR); 4949 CXRs between May 2020 and February 2021), CXRs from four Indonesia Awal Bros hospitals (three computed radiography (CR) and one DR; 1961 CXRs between September 2020 and October 2020), and two publicly available CXRs (Shenzhen and Montgomery). All acquired CXRs were labelled as TB or normal by a radiologist with 20-years’ experience.
[DAIP] First, exposure of CXRs were corrected by histogram normalization. To correct SID, lung regions of CXRs were cropped. For robustness of X-ray scattering (e.g., blurring), high or low frequency filters were applied to CXRs. Contrast of CXRs was changed using gamma correction for robustness of X-ray voltage variation. By applying above procedures, we obtained “generalized” dataset that covers diversity of CXRs.
[Experiments] Two networks (EfficientNet-B0), with and without applying DAIP pipeline, were trained using CXRs from Vietnam hospital. Classification performances were evaluated for all datasets.
Results
For dataset from Vietnam hospital, performances were almost same before and after applying pipeline (see Figure). For unseen datasets, performances consistently improved at similar level to Vietnam dataset (mean AUC gain: 0.06; p < 0.0001). In particular, one dataset (Indonesia 4) having substantially different image characteristics showed dramatic performance gain (AUC: 0.56 to 0.94). Our results indicate DAIP successfully generalized characteristics of CXRs.