Deep learning approach for automated cancer detection and tumor proportion score estimation of PD-L1 expression in lung adenocarcinoma

Background: This study proposed a computational method to detect the cancer areas and calculate the tumor proportion score (TPS) of PD-L1 immunohistochemistry (IHC) expression for lung adenocarcinoma based on deep learning and transfer learning. Patients and methods: PD-L1 22C3 and SP142 IHC slides of lung adenocarcinoma samples on digitized whole-slide images (WSI) database were employed. We build a deep convolutional neural network (DCNN) to automatically segment cancer regions. TPS was calculated based on segmented areas and then compared with the interpretations of pathologists. Results: We trained a DCNN model based on 22C3 dataset and fine-tuned it with SP142 dataset. We obtain a robust performance on cancer region detection on both datasets, with a sensitivity of 93.36% (22C3) and 92.80% (SP142) and a specificity of 93.97% (22C3) and 89.25% (SP142). With all the coefficient of determinations larger than 0.9 and Fleiss9 and Cohen9s Kappa larger than 0.8 (between mean or median of pathologists and TPS calculated by our method), we also found out the strong correlation between the TPS estimated by our computational method and estimation from multiple pathologists9 interpretations of 22C3 and SP142 respectively. Conclusion: We provide an AI method to efficiently predict cancer region and calculate TPS in PD-L1 IHC slide of lung adenocarcinoma on two different antibodies. It demonstrates the potential of using deep learning methods to conveniently access PD-L1 IHC status. In the future, we will further validate the AI tool for automated scoring PD-L1 in large volume samples.