Published
RSNA Annual Meeting (2023)
Authors
Heejun Shin1, Taehee Kim1, Hruthvik Raj1, Muhammad Shahid Jabbar1, Desalgn Abebaw Zeleke1, 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.
Purpose
Many AI methods to detect chest X-ray (CXR) abnormalities have demonstrated promising results by adopting image pre-processing techniques (e.g., histogram equalization (HE), contrast limited adaptive histogram equalization (CLAHE), and unsharp masking (UM)). However, those methods showed limited diagnostic performance when applied to CXRs with different image characteristics from various X-ray scanners. Here, we propose an X-ray physics-based data augmentation (i.e., XPA) that perturbates CXRs during AI training to overcome this problem.
Materials and Methods
Unlike conventional image pre-processing methods (e.g., HE, CLAHE, and UM) that normalize CXRs before AI training and testing, XPA randomly perturbates image characteristics on training CXRs by applying a series of image processing methods (e.g., gamma correction for contrast perturbation) to mimic hardware-related changes (e.g., voltage, current, etc.) during AI training.
Seven datasets from different X-ray machines (digital radiography (DR) or computed radiography (CR)) and institutions were collected. One dataset was from a Vietnam hospital and annotated by a radiologist as normal or abnormal (e.g., opacity, etc.) for AI (7,202 CXRs for training; 1,278 CXRs for testing (VHDR1)). Four datasets were from Indonesian hospitals, including a dataset acquired from a portable X-ray machine (IHCR1,Portable:204; IHCR2 :227; IHCR3:356; IHDR2:1,909 CXRs) and annotated for AI testing. Two datasets (i.e., Shenzhen (SZDR3) and Montgomery (MGCR4)) were from public domains as testing data.
We trained four AI models (EfficientNet-B6) using HE, CLAHE, UM, and proposed XPA to classify CXRs as normal or abnormal and compared their diagnostic performance. To check the capability of each method to cover CXRs from different machines, only the CXRs acquired using the DR system in the Vietnam hospital were utilized for training.
Results
For VHDR1 (internal test dataset), the diagnostic performance of all AI models was not statistically significant (i.e., p-value > 0.05). However, in most of the test datasets, the AI model with XPA outperformed the others, including the datasets acquired from the different CR detectors (IHCR2, IHCR3, and MGCR4), portable X-ray machine (IHCR1,Portable; AUC: 0.950 for XPA; 0.924 for HE; 0.920 for CLAHE; 0.891 for UM; p-value < 0.05), and DR detectors (IHDR2 and SZDR3).
Conclusions
The diagnostic performance of the AI model was improved with XPA for the CXR datasets from different X-ray machines (i.e., DR, CR, and portable machines) compared to those of the other AI models with conventional image pre-processing methods.
Clinical Relevance/Application
The proposed AI showed potential clinical usage when the CXRs were acquired using various X-ray scanners.