THREE-STATE CLASSIFICATION OF PULMONARY ARTERY WEDGE PRESSURE FROM CHEST X-RAY IMAGES USING CONVOLUTIONAL NEURAL NETWORKS

The pulmonary artery wedge pressure (PAWP) is an index used to evaluate pulmonary congestion caused by heart failure. In a previous study, a convolutional neural network (CNN) was used to estimate PAWP from chest X-ray images in binary states. This study is beneficial for medicine; however, there is a need to estimate PAWP in more detail. Therefore, we developed a CNN that outputs three classes depending on the PAWP (normal class: less than 12 mmHg; anomaly1 class: between 12 and 18 mmHg; anomaly2 class: 18 mmHg or more). The experiment used data of 936 patients, which were divided into training (80%) and test data (20%). Moreover, a validation dataset (20%) was extracted from the training dataset to tune the hyperparameters (learning rate and number of epochs). As a result of learning the CNNs, the optimal learning rate and epochs were 10^−5.5 and 96, respectively. The accuracy of the test data was approximately 63%. The accuracy of the normal class was sufficient; however, those of anomaly1 and anomaly2 classes were insufficient. Therefore, the estimation accuracy must be improved in future work.