DeepRePath: Finding prognostic features of tumor cell on the histopathologic image of lung cancer using explainable deep convolutional neural network

Abstract

Introduction: Histopathological images in high-powered contain quantitative information for tumor cell morphology, but humans could not sufficiently extract prognostic features of tumor cell morphology associated with tumor recurrence of patient. In this study, we developed and trained the DeepRePath model, a deep convolutional neural network (CNN), on histopathological images to predict the recurrence of resected lung adenocarcinoma (LUAD). Methods: A total of 244 haematoxylin and eosin stained histopathology slides of resected LUAD patients were used to train and validate the DeepRePath model. We manually segmented a cancer region in a LUAD tissue slide and captured to an image. All images were captured at x400 magnification and split into 13,329 image patches. We trained the DeepRePath model by transfer learning using 13,329 image patches from 244 LUAD patients. Results: 95 out of 244 patients developed recurrence after surgical resection. The DeepRePath model showed an average area under the curve (AUC) of 0.78 in 5-fold cross validation on the training cohort. We plan to employ external validation to evaluate the proposed model. In addition, using gradient-weighted class activation mapping (Grad-CAM), we confirmed that DeepRePath identified and used individual tumor cells with atypia to predict the recurrence of the patient. Conclusions: DeepRePath model using transfer learning and CNN can accurately predict recurrence after curative resection of LUAD based on atypia of tumor cell.