Classification of collagen remodeling in asthma using second-harmonic generation imaging, supervised machine learning and texture-based analysis.

Abstract

Airway remodeling is present in all stages of asthma severity and has been linked to reduced lung function, airway hyperresponsiveness and increased deposition of fibrillar collagens. Traditional histological staining methods used to visualize the fibrotic response are poorly suited to capture the morphological traits of extracellular matrix (ECM) proteins in their native state, hindering our understanding of disease pathology. Conversely, second harmonic generation (SHG), provides label-free, high-resolution visualization of fibrillar collagen; a primary ECM protein contributing to the loss of asthmatic lung elasticity. From a cohort of 13 human lung donors, SHG-imaged collagen belonging to non-asthmatic (control) and asthmatic donors was evaluated through a custom textural classification pipeline. Integrated with supervised machine learning, the pipeline enables the precise quantification and characterization of collagen, delineating amongst control and remodeled airways. Collagen distribution is quantified and characterized using 80 textural features belonging to the Gray Level Cooccurrence Matrix (GLCM), Gray Level Size Zone Matrix (GLSZM), Gray Level Run Length Matrix (GLRLM), Gray Level Dependence Matrix (GLDM) and Neighboring Gray Tone Difference Matrix (NGTDM). To denote an accurate subset of features reflective of fibrillar collagen formation; filter, wrapper, embedded and novel statistical methods were applied as feature refinement. Textural feature subsets of high predictor importance trained a support vector machine model, achieving an AUC-ROC of 94% ± 0.0001 in the classification of remodeled airway collagen vs. control lung tissue. Combined with detailed texture analysis and supervised ML, we demonstrate that morphological variation amongst remodeled SHG-imaged collagen in lung tissue can be successfully characterized.