Estimation of pulmonary function from time-resolved dynamic chest radiography using machine learning in patients with
Takehiro Shiinoki1, Yuki Yuasa2, Tsunahiko Hirano3
1Department of Radiation Oncology, Graduate School of Medicine, Yamaguchi University, Ube, Yamaguchi, Japan.
Background:
Pulmonary function tests (PFTs), particularly spirometry, are the reference standard for assessing airflow limitation in respiratory diseases such as chronic obstructive pulmonary disease (COPD) and interstitial pulmonary disease. However, spirometry requires substantial patient cooperation and may be unreliable in children, the elderly, and patients with cognitive impairment, and its use was further limited during the COVID-19 pandemic. Dynamic chest radiography (DCR), which captures sequential thoracic images during respiration at low radiation dose, has emerged as a promising modality for evaluating respiratory dynamics, but its potential to quantitatively estimate pulmonary function through radiomic analysis remains insufficiently explored.
Purpose:
This study aimed to determine the potential of radiomic features of the lung on DCR to predict pulmonary function (FEV1, forced expiratory volume in the first second; FVC, forced vital capacity) and to classify patients at high risk (FEV1/FVC).
Methods:
We retrospectively analysed data from 151 patients. The DCRs at end-inspiration (EI), end-expiration (EE), and the respiratory phase of maximum variation in the lung area from EI to EE (insp2expvmax) or from EE to EI (exp2inspvmax) were defined based on the lung area. A respiratory motion map was also calculated. To combine the defined DCR and respiratory motion map, feature extraction was performed, followed by the least absolute shrinkage and selection operator (LASSO). Predictive regression and classification models with various radiomic feature combinations (nos. 1-6) were constructed for pulmonary function. Pearson's correlation coefficients (R) were calculated for FEV1 and FVC, and the area under the curve (AUC) was calculated for FEV1/FVC. Our predictive models were compared using the conventional formula.
Results:
We constructed a predictive regression and classification model for FEV1, FVC, and FEV1/FVC ratio using DCR images and a respiratory motion map. The model accuracy with DCR at each respiratory phase and the respiratory motion map-based radiomic features was better than that of the conventional method.
Conclusions:
In this single-center retrospective study, radiomic features extracted from DCR at multiple respiratory phases combined with respiratory motion maps showed promise for estimating pulmonary function, outperforming conventional demographic-based prediction. External validation in multi-center cohorts is warranted before clinical translation.
