Clinically applicable histopathological diagnosis system for gastric cancer detection using deep learning

Abstract

The early detection and accurate histopathological diagnosis of gastric cancer increase the chances of successful treatment. The worldwide shortage of pathologists offers a unique opportunity for the use of artificial intelligence assistance systems to alleviate the workload and increase diagnostic accuracy. Here, we report a clinically applicable system developed at the Chinese PLA General Hospital, China, using a deep convolutional neural network trained with 2,123 pixel-level annotated H&E-stained whole slide images. The model achieves a sensitivity near 100% and an average specificity of 80.6% on a real-world test dataset with 3,212 whole slide images digitalized by three scanners. We show that the system could aid pathologists in improving diagnostic accuracy and preventing misdiagnoses. Moreover, we demonstrate that our system performs robustly with 1,582 whole slide images from two other medical centres. Our study suggests the feasibility and benefits of using histopathological artificial intelligence assistance systems in routine practice scenarios.

Fig. 1. The framework of our research and model performance on the daily gastric dataset.

a Deep learning model training and inference. We trained the model using WSIs digitalized and annotated at PLAGH. We illustrated the training data distribution at the slide level. The abbreviations are detailed in Supplementary Table 1. The trained model was tested by slides collected from PLAGH and two other hospitals. b The plot of the model performance histogram of the slides from the daily gastric dataset. c Model performance histogram of the daily gastric slides digitalized by three different scanners.

Fig. 2. Highlights of the deep learning model.

a Two cases detected by the AI assistance system that were initially misdiagnosed by pathologists. b Violin plot of the probability distributions for the malignant and benign cases in the IHC dataset. c Eight examples of false negative and false positive cases. The experiment was performed five times, and we obtained the same results.

Fig. 3. Experimental settings and examination results of the performance of the pathology trainees.

a Trainee pathologists were divided into 3 groups to make diagnoses on 100 slides of class I–VI. b The model prediction ROC curve and 12 pathologists' performance in the examination. c The average diagnostic accuracy of the three groups under different time settings. d Diagnostic consistency among different groups.

Fig. 4. Model performance on the multicentre dataset.

a The AUC, accuracy, sensitivity, specificity of the deep learning model on data collected from three hospitals. b ROC curves of the model on the multicentre dataset.