Machine-independent ai for chest x-ray abnormality classification

OBJECTIVE Many AI methods to detect chest X-ray (CXR) abnormalities have been developed using CXRs from a single machine (e.g., digital radiography (DR) system), reporting reduced diagnostic performance on CXRs from other machines (e.g., computed radiography (CR) systems). Here, we propose a machine-independent AI to address this problem. MATERIALS & METHODS 8,480 CXRs from a Vietnam hospital were acquired using a DR system (DR1) and annotated by a radiologist as normal or abnormal (e.g., opacity, etc.). 2,696 CXRs from four Indonesian hospitals (IHs) were acquired using different DR or CR systems (DR2 for IH1, CR1 for IH2, CR2 for IH3, CR3 for IH4) and annotated. We trained two AIs (baseline and proposed) using the CXRs from the Vietnam hospital (7,202 for training; 1,278 for testing) to classify CXRs as normal or abnormal. The baseline model was trained by utilizing the conventional CLAHE method. In contrast, the proposed model was trained by perturbating training data based on X-ray hardware-related changes (e.g., sharpness, contrast, and noise). RESULTS When we tested both AIs on CXRs from DR1, the diagnostic performance (AUC) was not different (0.962 w/ proposed; 0.962 w/ conventional; p=0.33). For the other CXRs from different machines, the proposed AI outperformed the conventional (DR2: 0.933 w/ proposed, 0.927 w/ conventional, p =0.07; CR1: 0.950 w/ proposed, 0.920 w/ conventional, p=0.004; CR2: 0.936 w/ proposed, 0.910 w/ conventional, p=0.012; CR3: 0.937 w/ proposed, 0.806 w/ conventional, p=0.007). CONCLUSION The proposed AI achieved good diagnostic performance (AUC > 0.93) over the CXRs from the different X-ray machines. Many AI methods to detect chest X-ray (CXR) abnormalities have been developed using CXRs from a single machine (e.g., digital radiography (DR) system), reporting reduced diagnostic performance on CXRs from other machines (e.g., computed radiography (CR) systems). Here, we propose a machine-independent AI to address this problem. 8,480 CXRs from a Vietnam hospital were acquired using a DR system (DR1) and annotated by a radiologist as normal or abnormal (e.g., opacity, etc.). 2,696 CXRs from four Indonesian hospitals (IHs) were acquired using different DR or CR systems (DR2 for IH1, CR1 for IH2, CR2 for IH3, CR3 for IH4) and annotated. We trained two AIs (baseline and proposed) using the CXRs from the Vietnam hospital (7,202 for training; 1,278 for testing) to classify CXRs as normal or abnormal. The baseline model was trained by utilizing the conventional CLAHE method. In contrast, the proposed model was trained by perturbating training data based on X-ray hardware-related changes (e.g., sharpness, contrast, and noise). When we tested both AIs on CXRs from DR1, the diagnostic performance (AUC) was not different (0.962 w/ proposed; 0.962 w/ conventional; p=0.33). For the other CXRs from different machines, the proposed AI outperformed the conventional (DR2: 0.933 w/ proposed, 0.927 w/ conventional, p =0.07; CR1: 0.950 w/ proposed, 0.920 w/ conventional, p=0.004; CR2: 0.936 w/ proposed, 0.910 w/ conventional, p=0.012; CR3: 0.937 w/ proposed, 0.806 w/ conventional, p=0.007). The proposed AI achieved good diagnostic performance (AUC > 0.93) over the CXRs from the different X-ray machines.