Two-step artificial intelligence system for endoscopic gastric biopsy improves the diagnostic accuracy of pathologists

BackgroundEndoscopic biopsy is the pivotal procedure for the diagnosis of gastric cancer. In this study, we applied whole-slide images (WSIs) of endoscopic gastric biopsy specimens to develop an endoscopic gastric biopsy assistant system (EGBAS). MethodsThe EGBAS was trained using 2373 WSIs expertly annotated and internally validated on 245 WSIs. A large-scale, multicenter test dataset of 2003 WSIs was used to externally evaluate EGBAS. Eight pathologists were compared with the EGBAS using a man-machine comparison test dataset. The fully manual performance of the pathologists was also compared with semi-manual performance using EGBAS assistance. ResultsThe average area under the curve of the EGBAS was 0 center dot 979 (0 center dot 958-0 center dot 990). For the diagnosis of all four categories, the overall accuracy of EGBAS was 86 center dot 95%, which was significantly higher than pathologists (P< 0 center dot 05). The EGBAS achieved a higher kappa score (0 center dot 880, very good kappa) than junior and senior pathologists (0 center dot 641 +/- 0 center dot 088 and 0 center dot 729 +/- 0 center dot 056). With EGBAS assistance, the overall accuracy (four-tier classification) of the pathologists increased from 66 center dot 49 +/- 7 center dot 73% to 73 center dot 83 +/- 5 center dot 73% (P< 0 center dot 05). The length of time for pathologists to manually complete the dataset was 461 center dot 44 +/- 117 center dot 96 minutes; this time was reduced to 305 center dot 71 +/- 82 center dot 43 minutes with EGBAS assistance (P = 0 center dot 00). ConclusionsThe EGBAS is a promising system for improving the diagnosis ability and reducing the workload of pathologists.