| 本文已被：浏览 1099次 下载 865次 |
码上扫一扫！ |
| 基于级联区域卷积神经网络算法在肾组织病理切片中对肾小球的识别与定位 |
| 杨会1,刘雪宇2,张兴娜1,姜秋竹1,刘云霄3,王晨4,李明2,李荣山1,周晓霜1* |
|
|
(1. 山西医科大学附属人民医院肾内科, 太原 030000;
2. 太原理工大学大数据学院, 太原 030024;
3. 山西医科大学附属人民医院病理科, 太原 030000;
4. 山西医科大学第二医院病理科, 太原 030000
*通信作者) |
|
| 摘要: |
| 目的 基于级联区域卷积神经网络（cascade R-CNN）算法开发一种能自动识别肾组织病理切片图像中肾小球的人工智能（AI）系统，帮助病理医师提高计算肾小球个数与识别肾小球的效率。方法 收集2017-2019年3年内在山西医科大学第二医院和山西医科大学附属人民医院行肾穿刺活检术患者的肾脏病理切片，剔除模糊不清、染色质量差的图像，最终得到1 180张质量无明显差异的六胺银（PASM）染色图像。通过高分辨率全视野数字切片（WSI）获得数字化扫描图像，图像数据通过远程病理系统传输到云端并储存。使用cascade R-CNN方法创建训练集（940张图像）和测试集（240张图像），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cascade R-CNN算法识别出肾小球的精确度和召回率。将测试集的病理切片由3名工作年限至少3年的病理医师阅读，计算医师们识别肾小球的精确度与时间。结果 基于cascade R-CNN网络训练完成的深度学习模型识别每张图像肾小球区域时间为（0.20±0.02）s。精确度、召回率分别为93.90%、98.00%，F1值为95.91%。3名病理医师识别每张图像肾小球区域时间分别为（3.57±0.05）、（4.57±0.07）、（3.98±0.02）s，精确度分别为88.08%、89.69%、89.98%，差异均无统计学意义（P均>0.05）。cascade R-CNN算法识别肾小球的精确度高于3名病理医师的平均精确度（89.25%），差异有统计学意义（t=-5.607，P=0.009）。结论 cascade R-CNN算法通过高分辨率WSI可快速有效地识别肾小球，能够帮助病理医师提高肾脏疾病的诊断效率。 |
| 关键词: 肾疾病 病理学 深度学习 卷积神经网络 肾小球 图像识别 |
| DOI：10.16781/j.0258-879x.2021.04.0445 |
| 投稿时间：2020-07-30修订日期：2020-12-18 |
| 基金项目:山西省重点研发计划项目（201803D31151），山西省基础研究项目（2015011098）. |
|
| Identification and localization of glomerulus in renal pathological sections based on cascade region-convolutional neural network algorithm |
| YANG Hui1,LIU Xue-yu2,ZHANG Xing-na1,JIANG Qiu-zhu1,LIU Yun-xiao3,WANG Chen4,LI Ming2,LI Rong-shan1,ZHOU Xiao-shuang1* |
(1. Department of Nephrology, People's Hospital Affiliated to Shanxi Medical University, Taiyuan 030000, Shanxi, China;
2. College of Big Data, Taiyuan University of Technology, Taiyuan 030024, Shanxi, China;
3. Department of Pathology, People's Hospital Affiliated to Shanxi Medical University, Taiyuan 030000, Shanxi, China;
4. Department of Pathology, the Second Hospital of Shanxi Medical University, Taiyuan 030000, Shanxi, China
*Corresponding author) |
| Abstract: |
| Objective To develop an artificial intelligence (AI) system that could automatically identify the glomerulus on the renal pathological section images based on cascade region-convolutional neural network (cascade R-CNN) algorithm, and to help pathologists to calculate the number of glomerulus and identify glomerulus. Methods Renal pathological sections from patients undergoing renal biopsy in the Second Hospital of Shanxi Medical University and People's Hospital Affiliated to Shanxi Medical University from 2017 to 2019 were collected. Totally, 1 180 periodic acid-silver metheramine (PASM) stained section images of similar quality were included after eliminating the blurred and poor quality ones. The digital scanned images were obtained by high-resolution whole slide image (WSI), and the image data were transmitted and stored to the cloud through the remote pathology system. A training set (940 images) and a test set (240 images) were created by cascade R-CNN. The training set was used to train AI to identify the glomerulus, and the test set was used to test and evaluate the precision and recall rate of cascade R-CNN algorithm to identify the glomerulus. The pathological sections of the test set were read by 3 pathologists who had worked for at least 3 years, and the precision and time for the pathologists to identify the glomerulus were calculated. Results The identification time of each glomerular region imnage was (0.20±0.02) s using the deep learning model trained by cascade R-CNN network. The precision and recall rate were 93.90% and 98.00%, respectively, and the F1 value was 95.91%. The time for the 3 pathologists to identify each glomerular region image was (3.57±0.05), (4.57±0.07), and (3.98±0.02) s, and the precision was 88.08%, 89.69%, and 89.98%, respectively, with no significant difference (all P>0.05). The precision of the cascade R-CNN algorithm was significantly higher than those of the 3 pathologists (89.25%) (t=-5.607, P=0.009). Conclusion cascade R-CNN algorithm can quickly and effectively identify glomerulus through high-resolution WSI, and it can help pathologists improving the diagnosis efficiency of renal diseases. |
| Key words: kidney diseases pathology deep learning convolutional neural network kidney glomerulus image recognition |
