Deep Learning-Based Segmentation of Lung Adenocarcinoma Whole-Slide Images for Objective Grading, Tumor Spread...

Manual quantification of morphologic patterns in lung adenocarcinoma is subject to reproducibility issues due to interpathologist variability. In this study, we developed a deep learning-based multiclass segmentation model providing a modality for objective and quantitative grading of digitized lung adenocarcinoma images from resected specimens. Additionally, the model can detect tumor spread through air spaces and show enrichment of specific morphologic patterns in tumors with different genomic alterations. The study was based on 766 resected nonmucinous lung adenocarcinomas. Deep Multi-Magnification Network was trained to segment 14 tissue subtypes based on annotations of 108 internal whole-slide images at pixel level by thoracic pathologists (J.C.C. and W.D.T.). The trained model was validated on an external cohort of 130 cases for determining predominant patterns and on the remaining 528 internal cases for the 3 clinical tasks. The model graded nonmucinous lung adenocarcinomas based on the International Association for the Study of Lung Cancer Pathology Committee recommendation and successfully stratified patients into well, moderately, and poorly differentiated morphologies (P < 1 × 10^-4). Pixels categorized as spread through air spaces significantly correlated with pathologists' interpretations. For molecular analysis, solid pattern was enriched with TP53 mutations and depleted of EGFR kinase domain mutations. Lepidic pattern was inversely associated with TP53 mutations. Acinar was enriched with EGFR mutations, whereas papillary was associated with RET fusions. Our study demonstrated that deep learning-based segmentation can accurately quantify histologic patterns in lung adenocarcinoma and identify additional prognostic features. By simultaneously providing an objective assessment of various tasks, our comprehensive methodology in lung adenocarcinoma paves way for deep learning-assisted pathologic diagnosis and treatment guidance.