Deep learning‐based lung volume estimation with dynamic chest radiography

TL;DR

This paper shows that deep learning models can accurately estimate lung volume from chest X-rays better than traditional methods.

Contribution

The novel contribution is applying deep learning to dynamic chest radiography for improved lung volume estimation.

Findings

VGG19 and DenseNet121 models outperformed linear regression in lung volume estimation with lower MAE and MAPE.

Estimated forced vital capacity (FVC) showed moderate correlation but higher error rates compared to reference values.

Abstract

Dynamic chest radiography (DCR) is a recently developed low‐dose pulmonary functional imaging method that can be performed in a general X‐ray room. DCR provides sequential images during respiration, and the measured changes in lung area are a promising diagnostic indicator of lung function.

To investigate lung volume estimation using deep learning from DCR images during respiration and evaluate its accuracy in comparison with previously proposed estimation methods.

Two convolutional neural networks (CNNs), VGG19 and DenseNet121, were trained using DCR image datasets from 257 patients, with reference lung volumes derived from corresponding computed tomography (CT) images. The performance of the models was evaluated using mean absolute error (MAE) and mean absolute percentage error (MAPE), and compared against that of a conventional linear regression model. Correlation between the estimated and reference lung volumes was assessed using Pearson's correlation coefficient (r) and the degrees‐of‐freedom‐adjusted coefficient of determination (Rf2
). Forced vital capacity (FVC) was also estimated by subtracting the lung volume at maximum exhalation from that at maximum inhalation.

The VGG19 and DenseNet121 models demonstrated superior performance in estimating whole lung volume (combined right and left lung) compared to the linear regression method. Specifically, MAE was 373/376 mL, MAPE was 8.1%/7.9%, r was 0.88/0.90, and Rf2
was 0.76/0.80 for VGG19/DenseNet121, respectively. In contrast, the linear regression model yielded an MAE of 568 mL, MAPE of 12.4%, r of 0.84, and Rf2
of 0.69. Although the Rf2
values for DCR‐derived FVC using VGG19 and DenseNet121 indicated moderate correlation, the MAE and MAPE were relatively high at 1.3/1.4 L and 41.1%/47.0%, respectively.

The proposed deep learning‐based approach for lung volume estimation from DCR images outperformed the conventional linear regression method. Further improvements in CNN model architecture and the incorporation of guided forced respiratory maneuvers may enhance the potential for image‐based pulmonary function testing.

Figures