Automatic Detection Of Early Gastric Cancer In Endoscopy Based On Mask Region-Based Convolutional Neural Networks(Mask R-CNN)

Background Gastric cancer(GC)was ranked 5th in incidence and 4th in mortality worldwide.Early gastric cancer(EGC)can be well treated by endoscopic resection,with a 5-year survival rate of 95%,much higher than that of advanced gastric cancer.Therefore,early detection,diagnosis,and treatment of GC were important measures to reduce GC mortality.Endoscopy was an important method for GC detection.Mucosal changes of EGC in white light imaging(WLI)were not obvious and often lead to misdiagnosis.More advanced image enhancement endoscopy such as narrow band imaging(NBI)required equipment and support from experienced doctors.Artificial intelligence(AI)was widely used in the medical field with the rapid development of computer technology.Computer vision combined with medical imaging has become an important direction,and AI has proven its potential for the diagnosis and analysis of EGC.Compared with traditional machine learning,deep learning was much more effective in extracting image features.As the representative branch of deep learning,the convolutional neural network(CNN)had excellent detection ability.CNN also had the limitation of large parameters,low efficiency and high demand for hardware.Mask R-CNN was developed based on Faster R-CNN and added the Mask branch.It gave attention to the accuracy,precision,and speed of image recognition,and has proven its excellent detection performance in many fields.Objective To develop an automatic endoscopic detection system based on Mask R-CNN for detecting EGC in WLI and NBI.Testing the performance of Mask R-CNN system on image and video datasets,and comparing the performance between Mask R-CNN and endoscopists.Evaluate the potential of Mask R-CNN for clinical application.Methods First,7869 WLIs and 3852 NBIs were obtained from the First Affiliated Hospital of Anhui Medical University and divided into the training set,validation set,image test set,and endoscopists test set according to a certain ratio.Comparing the performance of Mask R-CNN on the static test set with pathology results or endoscopists’ diagnosis.In addition,several videos were prospectively obtained in real time for speed measurement,until 10 WLI videos and 10 NBI videos containing EGC were included in the video test set.These 20 videos were used to test the dynamic performance of Mask R-CNN in EGC detection.Diagnostic ability was assessed in terms of accuracy,sensitivity,specificity,positive predictive value(PPV),and negative predictive value(NPV).Results The accuracy of the Mask R-CNN system in diagnosing EGC in WLIs was89.90%,and the sensitivity was 92.00%,the specificity was 89.20%.The accuracy of the diagnosis of EGC in NBIs was 84.80%,with a sensitivity of 98.25%,and a specificity of 81.00%.The Mask R-CNN results were consistent with the pathological diagnosis results in two modes.Mask R-CNN was able to capture real-time videos up to35 frames per second and detect EGC.The accuracy of diagnosing EGC in WLI videos was 89.21%,while the accuracy in NBI videos was 84.87%.In controlled experiments,the overall accuracy rate and the specificity of the Mask R-CNN system detecting EGC in WLI were 90.25% and 91.67%,higher than that of all endoscopists.The sensitivity of86.00% was similar to that of the experts.In NBI mode,the Mask R-CNN had an accuracy of 87.00% and a sensitivity of 96.00%,which were no less than the experts.The specificity of 78.00% was similar to that of all endoscopists.Conclusion Mask R-CNN had excellent performance in detecting EGC in still images and motion videos in both WLI and NBI modes.It made up for functional imperfections in relevant AI models at this stage and showed great potential for practical clinical applications.