Speed and efficiency analysis of artificial intelligence recognition system of glomerular pathology based on Faster R-CNN algorithm

Objective To develop an artificial intelligence (AI) system that can automatically identify the glomerulus in the images of renal histopathological sections, helping pathologists to improve the speed and efficiency of calculating the number of glomerulus and identifying ischaemia-sclerosing glomerulus. Methods A digital scan was performed for 11,476 kidney pathological sections stained by PASM from the People's Hospital of Shanxi Province and the Second Hospital of Shanxi Medical University from 2008 to 2018. The image data were transmitted to the cloud end and stored via a remote pathology system. The Faster R-CNN method was used to create a training set consisting of 2296 images and a test set consisting of 174 images. The training set was used to train AI to learn how to recognize glomeruli, and the test set was used to test and evaluate average time and accuracy of the recognition of glomeruli. At the same time, 174 pathological sections in the test set were read by 2 pathologists (junior doctors) who had worked for about 2 years and a pathologist (senior doctor) with more than 10 years of work experience, and the average time and accuracy of glomerular recognition by the pathologists. Results The model was obtained by training the AI system developed based on Faster R-CNN network. The performance of the AI model on the test set was:mAP=94.37%. The AI's processing time of the whole slide image was about 1 s, and the average time to recognize a glomerulus was ≤ 0.050±0.040 s (The data was provided by Taiyuan University of Technology). The times for junior pathologists and senior ones to recognize a glomerulus were (22±2.321) s and (11±1.425) s, and the recognition time was slower than AI (both P<0.05). The accuracy values of the glomerular identification of junior and senior physicians were (82±4.92)% and (93±7.64)%, and the accuracy of AI was (99.93±1.30)%. The accuracy of AI's glomerular recognition was superior to junior and senior pathologists' recognition (both P<0.05). Conclusions The speed and efficiency of recognizing ischemic sclerosing glomeruli and calculating the number of glomeruli based on the AI system developed by Faster R-CNN are significantly higher than those of pathologists participating in this study.