Multi-resolution classification of exhaled aerosol images to detect obstructive lung diseases in small airways

Exhaled aerosol patterns have been used to detect obstructive respiratory diseases in the upper airways. Signals from small airway diseases are weak and may not manifest themselves in the exhaled aerosol patterns. Therefore, it will be more challenging to detect abnormalities in small airways. The objective of this study is to develop a simulation-based classification model that can accurately classify small airway diseases. The model performance was evaluated in five obstructed models that are located in lung bifurcations G7-9. The exhaled aerosol images were quantified using local fractal dimensions at different sampling resolutions (n × n). The datasets were classified using both the random forest (RF) and support vector machine (SVM) algorithms. Results show that RF performs slightly and persistently better than SVM. The sampling resolution of 12 × 12 gave the optimal classification for both algorithms. Based on the lung models with predefined obstructive levels, the optimal classification accuracy is 87.0% for 5-class classification, and is 92.5% for 4-class classification by regrouping the mislabeled samples. The proposed model with multi-resolution fractal feature extraction and RF algorithm appears to be sensitive enough to accurately distinguish airway abnormalities in small airways beyond G7 with healthy bronchiole diameter <4 mm. This aerosol-based breath test is promising to develop into an alternative or supplemental tool to the low-dose CT scanning for lung cancer screening.