Tumor Cellularity Assessment Using Artificial Intelligence Trained on Immunohistochemistry-Restained Slides Improves Selection of Lung Adenocarcinoma Samples for Molecular Testing.

Abstract

Tumor cellularity (TC) in lung adenocarcinoma slides submitted for molecular testing is important in identifying actionable mutations, but absent best practice guidelines yield high interobserver variability in TC assessments. An artificial intelligence (AI)-based pipeline developed to assess TC in hematoxylin and eosin (H&E) whole slide images (WSIs) and in tumor areas (TAs) within WSIs includes a new model (CaBeSt-Net) trained to mask cancer cells, benign epithelial cells, and stroma in H&E WSIs using immunohistochemistry-restained slides, and a model to detect all cell nuclei. High masking accuracy (>91%) by CaBeSt-Net computed using 1024 H&E regions of interest and intraclass correlation coefficient >0.97 assessing TC assessments reliability by one pathologist and AI in 20 test regions of interest supported the pipeline's applicability to TC assessment in 50 study H&E WSIs. Using the pipeline, TCs assessed in TAs and WSIs were compared with those by three pathologists. Reliabilities of these ratings by the pathologists supported by the pipeline were good (intraclass correlation coefficient >0.82, P < 0.0001). The consistency of sample categorizations as inadequate or adequate (TC ≤ 20% cut point) for molecular testing among the pathologists assessing TCs without AI support was moderate in TAs (κ = 0.410, P < 0.0001) and slight in WSIs (κ = 0.132, nonsignificant). With AI support, the consistency was substantial in both WSIs (κ = 0.602, P < 0.0001) and TAs (κ = 0.704, P < 0.0001). By visualizing cancer and measuring TC in the sample, this novel AI-based pipeline assists pathologists in selecting samples for molecular testing.