Deep learning-based diagnosis of lung cancer using a nationwide respiratory cytology image set: improving accuracy and inter-observer variability

Abstract

Deep learning (DL)-based image analysis has recently seen widespread application in digital pathology. Recent studies utilizing DL in cytopathology have shown promising results, however, the development of DL models for respiratory specimens is limited. In this study, we designed a DL model to improve lung cancer diagnosis accuracy using cytological images from the respiratory tract. This retrospective, multicenter study used digital cytology images of respiratory specimens from a quality-controlled national dataset collected from over 200 institutions. The image processing involves generating extended z-stack images to reduce the phase difference of cell clusters, color normalizing, and cropping image patches to 256 x 256 pixels. The accuracy of diagnosing lung cancer in humans from image patches before and after receiving AI assistance was compared. 30,590 image patches (1,273 whole slide images [WSIs]) were divided into 27,362 (1,146 WSIs) for training, 2,928 (126 WSIs) for validation, and 1,272 (1,272 WSIs) for testing. The Densenet121 model, which showed the best performance among six convolutional neural network models, was used for analysis. The results of sensitivity, specificity, and accuracy were 95.9%, 98.2%, and 96.9% respectively, outperforming the average of three experienced pathologists. The accuracy of pathologists after receiving AI assistance improved from 82.9% to 95.9%, and the inter-rater agreement of Fleiss' Kappa value was improved from 0.553 to 0.908. In conclusion, this study demonstrated that a DL model was effective in diagnosing lung cancer in respiratory cytology. By increasing diagnostic accuracy and reducing inter-observer variability, AI has the potential to enhance the diagnostic capabilities of pathologists.