Impact of chest radiograph image size and augmentation on estimating pulmonary artery wedge pressure by regression convolutional neural network

• Main Authors: Omae, Yuto, Saito, Yuki, Fukamachi, Daisuke, Nagashima, Koichi, Okumura, Yasuo, Toyotani, Jun
• Format: Conference Proceeding
• Language: English
• Subjects: Artificial neural networks; Chest; Data augmentation; Heart diseases; Heart failure; Neural networks; Pulmonary arteries; Radiographs; Regression
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0162766

Heart failure is related to pulmonary artery wedge pressure (PAWP), which is one of the measurements for diagnosing heart disease. In the case of suspected heart failure, it is desirable to measure PAWP by right heart catheterization (RHC). However, RHC is an invasive procedure accompanied with the risk of complication. Therefore, a method to quantitatively estimate PAWP from chest radiographs by a regression convolutional neural network (R-CNN) was proposed as the previous study. The risk of complication is eliminated because the method is non-invasive. Moreover, developed R-CNN includes regression activation map (RAM), which is one of the white-box techniques for CNN. However, tuning hyper parameters of R-CNN (e.g., input image size and data augmentation) developed in previous researches, is insufficient. Therefore, we carry out sensitivity analyses of input image sizes and data augmentation against estimating PAWP from chest radiographs. Through these analyses, we found the appropriate input image size and data augmentation.