基于Faster R-CNN算法开发的肾小球病理人工智能识别系统的速度与效率分析

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

目的基于Faster R-CNN算法开发出能够自动对肾组织病理切片图像中肾小球进行识别的人工智能(artificial intelligence,AI)系统,帮助病理医师提高计算肾小球个数与识别缺血硬化性肾小球的速度和效率。方法将山西省人民医院和山西医科大学第二医院自2008年至2018年的11476例肾病患者PASM染色的肾脏病理切片进行数字化扫描,图像数据通过远程病理系统传输到云端并进行储存。使用Faster R-CNN方法创建包括2296张图像的训练集和包括174张图像的测试集,训练集用于训练AI学习识别肾小球,测试集用于测试和评价AI识别出肾小球的平均时间和准确率。同时将测试集的174张病理切片分别给工作2年左右的病理科医师(初级医师)和10年以上工作经历的病理科医师(高级医师)阅读,收集医师识别出肾小球的平均时间和准确率。结果通过训练基于Faster R-CNN网络开发的AI得到模型,AI模型在测试集上的性能为:mAP=94.37%。AI处理整张玻片图像处理时间约为1 s,平均识别一个肾小球的时间(0.05±0.04)s(数据由太原理工大学大数据库学院提供)。病理科初级医师和高级医师识别一个肾小球的时间为(22.32±2.32)s和(11.50±1.42)s,识别时间均慢于AI(均P<0.05)。初级医师和高级医师识别肾小球的精确度分别为(82.18±4.92)%和(93.29±7.64)%,AI为(99.93±1.30)%,AI识别肾小球的精确度优于初级医师和高级医师(均P<0.05)。结论基于Faster R-CNN方法开发的AI系统计算肾小球个数与识别缺血硬化性肾小球的速度和效率明显高于参与这项研究的病理科医师。

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.