Deep learning-based lung volume estimation with dynamic chest radiography.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Function Testing

Background:

• Dynamic Chest Radiography (DCR) is an emerging low-dose imaging technique for assessing lung function.
• DCR captures sequential X-ray images during respiration, offering insights into lung volume changes.

Purpose of the Study:

• To evaluate deep learning models for estimating lung volume from DCR images.
• To compare the accuracy of deep learning methods against conventional estimation techniques.

Main Methods:

• Two Convolutional Neural Networks (CNNs), VGG19 and DenseNet121, were trained on DCR datasets from 257 patients.
• Reference lung volumes were obtained from Computed Tomography (CT) scans; model performance was assessed using MAE, MAPE, and Pearson's correlation coefficient (r).

Main Results:

• VGG19 and DenseNet121 models significantly outperformed linear regression in whole lung volume estimation (e.g., VGG19: MAE 373 mL, r 0.88; DenseNet121: MAE 376 mL, r 0.90).
• Linear regression showed higher error (MAE 568 mL, r 0.84).
• DCR-derived Forced Vital Capacity (FVC) estimation using CNNs showed moderate correlation but higher relative errors.

Conclusions:

• Deep learning approaches provide a more accurate method for lung volume estimation from DCR compared to linear regression.
• Future research focusing on CNN architecture and guided respiratory maneuvers could further enhance DCR's utility in pulmonary function testing.