| 本文已被：浏览 872次 下载 1177次 |
码上扫一扫！ |
| 基于人工智能算法和数字病理切片对非酒精性脂肪性肝病病理特征的识别效果 |
| 孙丹辉1,荣义辉1,廖心怡2,潘雅婷2,王珏3,黄萍3,朱思越2,柳苏桐2,王亚妮2,杜帆4,于观贞2,4* |
|
|
(1. 北京大学国际医院感染性疾病科，北京 102206;
2. 上海中医药大学附属龙华医院肿瘤科，上海 200032;
3. 上海中医药大学附属龙华医院脾胃病研究所，上海 200032;
4. 浙江数字内容研究院数字医疗与人工智能实验室，绍兴 312366
*通信作者) |
|
| 摘要: |
| 目的 开发基于人工智能算法的非酒精性脂肪性肝病（NAFLD）病理特征识别模型，探究模型能否识别并可视化脂肪变性细胞、炎症细胞和纤维化等病理特征。方法 选择65只NAFLD小鼠的肝组织H-E染色和天狼猩红染色病理切片各65张，通过数字化扫描获得数字病理切片。对于H-E染色切片，使用CaseViewer 2.3软件在放大200、300、400倍后截取病变部位图像各2张，共获得390张脂肪变性细胞病理图像和390张炎症细胞病理图像；将图像上传至Horizope标注平台进行手动标注，然后通过数据增强得到2 340张脂肪变性细胞图像及2 340张炎症细胞图像；按4∶1∶1划分为训练集、验证集和测试集，其中训练集（1 560张）、验证集（390张）用于U-Net深度学习模型的训练学习和参数迭代，测试集（390张）用于模型的识别分析。采用Dice相似系数（DSC）、平均交互比（MIoU）、平均准确度（MA）和灵敏度对模型性能进行评估。对于天狼猩红染色切片，使用CaseViewer 2.3软件在放大50倍后进行全视野截取，采用了颜色特征提取算法进行纤维化识别。对130张数字病理切片进行人工NAFLD活动度积分（NAS）评分和机器评分，并计算和分析脂肪变性细胞面积占比（PFA）、炎症细胞密度（DIC）和纤维化面积占比（RFA）。结果 基于人工智能算法的NAFLD病理特征识别模型识别脂肪变性细胞的DSC为0.87，MIoU为0.80，MA为0.88，灵敏度为0.84；识别炎症细胞的DSC为0.84，MIoU为0.78，MA为0.85，灵敏度为0.80。65张病理切片的PFA为0.371（0.013~0.743），DIC为288（19~894）/mm2，RFA为0.048 5±0.025 4，PFA、DIC、RFA均与机器评分和人工NAS评分呈正相关（rs＝0.953和0.928、0.883和0.869、0.887和0.749，P均＜0.001）。结论 基于人工智能算法的NAFLD病理特征识别模型有良好的表现，能够帮助病理医师识别NAFLD的病理特征、提高识别效率与准确率。 |
| 关键词: 非酒精性脂肪性肝病 脂肪变性 炎症细胞 纤维化 U-Net 深度学习 |
| DOI：10.16781/j.CN31-2187/R.20220304 |
| 投稿时间：2022-04-14修订日期：2022-07-18 |
| 基金项目:上海中医药大学附属龙华医院高层次人才引进科研启动经费(LH02.51.002) |
|
| Recognition of pathological features of non-alcoholic fatty liver disease based on artificial intelligence algorithm and digital pathological sections |
| SUN Dan-hui1,RONG Yi-hui1,LIAO Xin-yi2,PAN Ya-ting2,WANG Jue3,HUANG Ping3,ZHU Si-yue2,LIU Su-tong2,WANG Ya-ni2,DU Fan4,YU Guan-zhen2,4* |
(1. Department of Infectious Diseases, Peking University International Hospital, Beijing 102206, China;
2. Department of Oncology, Longhua Hospital Affiliated to Shanghai University of Traditional Chinese Medicine, Shanghai 200032, China;
3. Institute of Spleen and Stomach Diseases, Longhua Hospital Affiliated to Shanghai University of Traditional Chinese Medicine, Shanghai 200032, China;
4. Laboratory of Digital Health and Artificial Intelligence, Zhejiang Digital Content Research Institute, Shaoxing 312366, Zhejiang, China
*Corresponding author) |
| Abstract: |
| ObjectiveTo develop a recognition model for the pathological features of non-alcoholic fatty liver disease (NAFLD) based on artificial intelligence (AI) algorithm, and to explore whether the model can recognize and visualize the pathological features such as steatosis cells, inflammatory cells, and fibrosis.MethodsSixty-five hematoxylin-eosin (H-E) stained and 65 picrosirius red stained pathological sections of liver tissues of 65 NAFLD mice were selected, and digital pathological sections were obtained by digital scanning. For H-E stained sections, 2 images of the lesion site were taken after 200, 300 and 400 times magnification using CaseViewer 2.3 software, and 390 pathological images of steatosis cells and 390 pathological images of inflammatory cells were obtained; the images were uploaded to the Horizope annotation platform for manual annotation, and then 2 340 images of steatosis cells and 2 340 images of inflammatory cells were obtained after data enhancement; and they were divided into training set, validation set and test set according to 4∶1∶1, of which training set (1 560 images) and validation set (390 images) were used for learning training and parameter iteration of U-Net deep learning model, and test set (390 images) was used for identification and analysis. Dice's similarity coefficient (DSC), mean intersection over union (MIoU), mean accuracy (MA), and sensitivity were used to evaluate the performance of the model. For the picrosirius red stained sections, CaseViewer 2.3 software was used to perform full field interception after 50 times magnification, and color feature extraction algorithm was used to identify fibrosis. Artificial NAFLD activity score (NAS) and machine score were performed on 130 digital pathological sections, and the proportion of fatty degeneration cell area (PFA), density of inflammatory cell (DIC), and ratio of fibrotic area (RFA) were calculated and analyzed.ResultsThe DSC, MIoU, MA and sensitivity of the NAFLD pathological feature recognition model based on AI algorithm were 0.87, 0.80, 0.88 and 0.84 for identifying steatosis cells, respectively; and 0.84, 0.78, 0.85 and 0.80 for identifying inflammatory cells, respectively. The PFA, DIC and RFA of the 65 digital pathological sections were 0.371 (0.013-0.743), 288 (19-894)/mm2 and 0.048 5±0.025 4, respectively. PFA, DIC and RFA were all positively correlated with the machine score and artificial NAS score (rs＝0.953 and 0.928, 0.883 and 0.869, and 0.887 and 0.749, all P < 0.001).ConclusionNAFLD pathological feature recognition model based on AI algorithm has good performance, and it can help pathologists identifying pathological features, so as to improve recognition efficiency and accuracy. |
| Key words: non-alcoholic fatty liver disease steatosis inflammatory cells fibrosis U-Net deep learning |
