Auchincloss, Hugh

Abstract Introduction: Endobronchial optical coherence tomography (EB-OCT) imaging has unique potential to diagnose and quantify early, microscopic disease changes and assess therapeutic responsivity in interstitial lung disease (ILD). However, training and time requirements for interpretation and quantitative evaluation of large volumetric EB-OCT datasets need to be optimized to facilitate clinical adoption. We develop and validate a computationally-efficient, rapid EB-OCT quantitative image analysis framework using artificial intelligence (AI) with deep learning architecture for 2D feature segmentation, 3D feature burden and spatial distribution mapping, and computer-aided diagnosis in early ILD subjects. Methods: Volumetric EB-OCT datasets, consisting of 2D cross-sectional images, were acquired from multiple locations in the bilateral lungs of ILD and healthy control subjects. EB-OCT features were manually segmented from cross-sectional images by an expert EB-OCT reader, including fibrosis, microscopic honeycombing, traction bronchiolectasis, normal parenchyma, and emphysema. EB-OCT images with corresponding feature segmentations were used to train and validate a multiclass convolutional neural network. 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 subjects using dice similarity score, precision, recall/sensitivity, specificity, and balanced accuracy. To generate 3D AI-based quantitative feature burden and spatial distribution maps, AI-based feature segmentation was performed on each EB-OCT 2D cross-sectional image within a volumetric imaging site. Three external pathologist readers independently interpreted 3D AI-based features maps for each EB-OCT imaging site acquired for a subject (without access to EB-OCT images) and provided a single diagnosis of usual interstitial pneumonia (UIP) or non-UIP ILD. Performance of computer-aided ILD diagnosis was compared against independently evaluated histopathology, clinical diagnosis, and prior EB-OCT image interpretation. Results: EB-OCT was acquired in 51 subjects: 30 in vivo early ILD, 5 in vivo non-ILD controls, and 16 ex vivo end-stage ILD. Overall pixel similarity between 2D manual and AI-based segmentation demonstrated high accuracy for each ILD feature (mean balanced accuracy 0.9). 3D quantitative AI-based feature maps demonstrated strong agreement (Spearman ρ0.87, p<0.001) between manual and AI-based segmentation. AI-based computer-aided diagnosis demonstrated 100% sensitivity/specificity for histologic UIP and clinical IPF for all pathologist readers, with improved diagnostic accuracy and significantly reduced training (45-minute training, 75% reduction) and interpretation time (<1 minute per subject for computer-aided diagnosis, 88% reduction) compared to EB-OCT image interpretation. Conclusion: The AI-based EB-OCT image analysis framework could expedite computer-aided diagnosis of early ILD and robust, quantitative assessment of microscopic disease burden, progression, and therapeutic responsivity.