Automated estimation of total lung volume using chest radiographs and deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Total lung volume is a critical biomarker for assessing restrictive lung diseases.
• Accurate measurement of total lung volume is essential for clinical diagnosis and management.

Purpose of the Study:

• To investigate the performance of deep learning (DL) approaches for automated total lung volume measurement from chest radiographs (CXRs).
• To evaluate the accuracy and reliability of DL models in predicting total lung volume compared to reference standards.

Main Methods:

• Trained DL models on a dataset of 7621 CXRs with CT-derived lung volumes and 928 CXRs with pulmonary function test (PFT)-derived lung volumes.
• Utilized stepwise complexity experiments to assess the impact of training data (CT vs. PFT) on model performance.
• Evaluated optimal models on 291 independent CXR studies using MAE, MAPE, and Pearson's r.

Main Results:

• The optimal DL model achieved a Mean Absolute Error (MAE) of 408 ml and a Mean Absolute Percentage Error (MAPE) of 8.1% using both frontal and lateral CXRs.
• Model predictions showed high correlation with reference standards (Pearson's r = 0.92).
• Fine-tuning with PFT-derived labels after CT-derived pretraining yielded optimal performance.

Conclusions:

• State-of-the-art deep learning accurately measures total lung volume from plain chest radiographs.
• The publicly available model offers a cost-effective tool for routine lung volume assessment.
• This DL system can assist in monitoring lung volume changes in patients over time.