Accurate measurement of cardio-thoracic ratio for cardiomegaly detection on chest radiographs using ai

OBJECTIVE Recently, many AI methods have been proposed to detect cardiomegaly on chest X-rays (CXRs) by measuring cardio-thoracic ratios (CTRs). However, accurate measurement of CTRs on CXRs with pulmonary abnormalities is challenging due to the corrupted lung regions. Here, we propose a new AI model to address this problem. MATERIALS & METHODS 931 CXRs from Indonesian hospitals, confirmed as not having any chest abnormality (i.e., normal CXRs), were annotated by three radiologists who measured CTRs for each CXR based on the boundaries of heart and lung regions. These were later split for AI training and testing. 170 CXRs, confirmed as having at least one of four abnormalities (i.e., cardiomegaly, effusion, opacity, and pneumothorax), were collected from a Vietnam hospital and annotated.Two AI models (baseline and proposed) were trained and evaluated by calculating mean absolute errors (MAEs) of CTR measurement. The baseline network was trained using the normal CXRs without any pulmonary abnormalities. The proposed network was trained using the normal and synthetic CXRs with pulmonary abnormalities (e.g., effusion, opacity, etc.) which were generated from the normal CXRs via another AI. RESULTS When we tested both AIs on normal CXRs (from Indonesia), a marginal improvement of CTR measurement was observed in the proposed network (MAE: 0.013 for proposed; 0.014 for baseline; p-value=0.028). On the other hand, when testing on CXRs with abnormalities (from Vietnam), the proposed network outperformed the baseline (MAE: 0.022 for proposed; 0.026 for baseline; p-value=0.001). CONCLUSION The proposed AI can accurately measure CTR values on CXRs with pulmonary abnormalities. Recently, many AI methods have been proposed to detect cardiomegaly on chest X-rays (CXRs) by measuring cardio-thoracic ratios (CTRs). However, accurate measurement of CTRs on CXRs with pulmonary abnormalities is challenging due to the corrupted lung regions. Here, we propose a new AI model to address this problem. 931 CXRs from Indonesian hospitals, confirmed as not having any chest abnormality (i.e., normal CXRs), were annotated by three radiologists who measured CTRs for each CXR based on the boundaries of heart and lung regions. These were later split for AI training and testing. 170 CXRs, confirmed as having at least one of four abnormalities (i.e., cardiomegaly, effusion, opacity, and pneumothorax), were collected from a Vietnam hospital and annotated.Two AI models (baseline and proposed) were trained and evaluated by calculating mean absolute errors (MAEs) of CTR measurement. The baseline network was trained using the normal CXRs without any pulmonary abnormalities. The proposed network was trained using the normal and synthetic CXRs with pulmonary abnormalities (e.g., effusion, opacity, etc.) which were generated from the normal CXRs via another AI. When we tested both AIs on normal CXRs (from Indonesia), a marginal improvement of CTR measurement was observed in the proposed network (MAE: 0.013 for proposed; 0.014 for baseline; p-value=0.028). On the other hand, when testing on CXRs with abnormalities (from Vietnam), the proposed network outperformed the baseline (MAE: 0.022 for proposed; 0.026 for baseline; p-value=0.001). The proposed AI can accurately measure CTR values on CXRs with pulmonary abnormalities.