Abstract Introduction Early, objective measurement of microscopic disease change is essential for early intervention and therapeutic responsivity assessment in interstitial lung disease (ILD). Endobronchial optical coherence tomography (EB-OCT) can provide low-risk, in vivo microscopic ILD diagnosis and repeat assessment of disease changes over time. Full clinical adoption of EB-OCT requires overcoming the current time-and labor-intensive training and qualitative interpretation of large volumetric imaging datasets. We develop and validate a quantitative analysis framework using artificial intelligence (AI)-based EB-OCT for 2D ILD feature detection, 3D volumetric feature burden and spatial distribution mapping, and 4D quantification of microscopic disease change over time on repeat EB-OCT. Methods EB-OCT datasets of sequential 2D cross-sectional images across a volumetric imaging site were acquired from multiple locations in the bilateral lungs of in vivo ILD and healthy control subjects and ex vivo ILD pneumonectomy specimens. Salient EB-OCT features including destructive fibrosis, microscopic honeycombing, traction bronchiolectasis and preserved lung parenchyma were segmented from 2D EB-OCT images by an expert EB-OCT reader, and used to train and validate a multiclass deep learning algorithm. The 2D AI-based segmentation model was subsequently evaluated against manual ‘ground truth’ segmentations for each feature in a de novo ‘test’ dataset of independent in vivo ILD subjects. 3D AI-based quantitative feature burden and spatial distribution maps were generated by performing AI-based feature segmentation iteratively on each 2D cross-sectional image within a volumetric EB-OCT imaging site, and compared against manual segmentation-based quantification. To achieve 4D time-serial microscopic assessment, ILD subjects were imaged with EB-OCT imaging at two timepoints, ∼6 months apart. AI-based 3D quantification of changes in disease burden and distribution, including presence/absence and increase/decrease of ILD features, was performed across matched paired EB-OCT imaging sites. Results The 2D AI-based automated feature segmentation showed 90% accuracy compared to expert manual feature segmentation. The AI-based 3D volumetric feature quantification and distribution demonstrated high accuracy compared to manual feature segmentation and quantification with significantly faster segmentation time (100x faster). 4D AI-based EB-OCT demonstrated the potential to objectively measure microscopic changes in disease burden and distribution in matched paired 3D EB-OCT imaging sites over time. Conclusion We demonstrate the feasibility of a 4D quantitative EB-OCT image analysis framework that provides precise computer-aided, quantitative assessment of microscopic disease burden and distribution, and repeat objective measurement of microscopic changes over time. This could have major implications in early diagnosis, and reliable assessment of progression and therapeutic responsivity at the microscopic level. This abstract is funded by: NIH
Nandy et al. (Fri,) studied this question.