Estimating Lung Volume Capacity from X-ray Images Using Deep Learning

Abstract

Estimating lung volume capacity is crucial in clinical medicine, especially in disease diagnostics. However, the existing estimation methods are complex and expensive, which require experts to handle and consequently are more error-prone and time-consuming. Thus, developing an automatic measurement system without a human operator that is less prone to human error and, thus, more accurate has always been a prerequisite. The limitation of radiation dose and various medical conditions in technologies like computed tomography was also the primary concern in the past. Although qualitative prediction of lung volume may be a trivial task, designing clinically relevant and automated methods that effectively incorporate imaging data is a challenging problem. This paper proposes a novel multi-tasking-based automatic lung volume estimation method using deep learning that jointly learns segmentation and regression of volume estimation. The two networks, namely, segmentation and regression networks, are sequentially operated with some shared layers. The segmentation network segments the X-ray images, whose output is regressed by the regression network to determine the final lung volume. Besides, the dataset used in the proposed method is collected from three different secondary sources. The experimental results show that the proposed multi-tasking approach performs better than the individual networks. Further analysis of the multi-tasking approach with two different networks, namely, UNet and HRNet, shows that the network with HRNet performs better than the network with UNet with less volume estimation mean square error of 0.0010.