흉부 측면 X 선 영상에서 심혈관 경계의 딥러닝 기반 자동 분석 소프트웨어 개발 및 검증

Abstract

Cardiovascular border (CB) analysis is a fundamental method for detecting and assessing the severity of heart diseases using chest X-ray (CXR). This study aimed to develop and validate a deep learning-based CB automatic analysis software algorithm for quantitative analysis of lateral chest X-ray images. For detecting CB in Lateral CXR images, we utilized the Mask R- CNN. This model is an algorithm capable of accurate object detection and segmentation. We proposed more precise results using reasonable pre-processing and post-processing algorithms with this model, and suggested indicator for quantitative analysis of images. The CB detection performance of the developed model was evaluated with a precision of 99.92%, recall of 99.89%, F1 score of 99.91%, and a false positive rate per patient of 0.48. We used a developed validation dataset to verify the reliability between the CB automatic label (CB_auto) and the CB manual label (CB_hand) using the proposed CB measure index. At the 4cm point in the proposed CB measure, the highest Intraclass correlation coefficient (ICC) (0.95-0.99) was observed. Comparing the absolute difference in CB measure between the normal control group and abnormal groups at the 4cm point, significant differences were found in Tricuspid valve disease (TD) 12.7mm, Mitral valve disease (MD) 11.2mm, Pulmonary valve disease (PD) 8.5mm, and Aortic valve disease (AD) 6.8mm. The success rate of the CB measure was 91.6% in normal, 83.0% in AD, 84.5% in MD, 84.4% in PD, and 77.7% in TD. Subsequently, we conducted statistical analysis on the measured CB measure results by gender (male, female) in the normal control group. The Mean and Standard deviation in the normal control group were 25.5±9.7 for males and 16.7±8.9 for females, and the median (Interquartile ranges (IQRs)) were 25.5 (18.9-31.9) for males and 16.7 (10.7-22.6) for females. Additionally, we compared the median (IQRs) values of z-scores after fitting the model for the normal control group using the GAMLSS library. For males, normal was 0.01 (-0.69 to +0.68), AD was -0.58 (-1.51 to +0.29), MD was -0.90 (-1.62 to -0.10), PD was -0.17 (-1.03 to +0.21), and TD was -1.09 (- 1.78 to +0.30). For females, normal was 0.01 (-0.67 to +0.68), AD was -0.33 (-1.12 to +0.42), MD was -0.91 (-1.57 to -0.22), PD was -0.80 (-1.57 to +0.13), and TD was -0.90 (-1.47 to 0.00). When comparing normal control with abnormal groups, the abnormal groups generally showed lower values.