| 基于迁移学习算法开发的人工智能脑积水诊断模型的临床应用 |
| |
| 引用本文： | 陈宇航,林中松,李新瑜,张,亮,李小潘,郝晓伟,程小兵,张鸿日.基于迁移学习算法开发的人工智能脑积水诊断模型的临床应用[J].中国临床神经外科杂志,2021,26(5):349-351. |
| |
| 作者姓名： | 陈宇航 林中松 李新瑜 张 亮 李小潘 郝晓伟 程小兵 张鸿日 |
| |
| 作者单位： | 471000 河南洛阳，河南科技大学第一附属医院神经外科（陈宇航、郝晓伟、程小兵、张鸿日），影像中心（李新瑜、李小潘）；201203 上海，上海纳凝微信息科技有限公司（林中松、张 亮） |
| |
| 摘 要： | 目的 利用迁移学习算法设计、开发一种人工智能脑积水影像诊断工具，并评价其应用颅脑CT平扫影像诊断脑积水的效果。方法 收集河南科技大学第一附属医院正常成人及脑积水的颅脑CT影像DICOM数据各1 250例，按6∶2∶2随机分为训练集、验证集和测试集。应用Python开发标记工具对数据进行预处理并标注特征变量，建立自主研发的逐像素ForrestNet-CNN算法模型，并对特征变量进行提取和深度学习。选取影像学住院医师、主治医师、副主任医师各2名对250例脑积水和250例正常影像进行测试，并与人工智能模型的测试结果进行比较。结果 人工智能模型、住院医师、主治医师、副主任医师诊断脑积水的敏感性分别为94.4%、92.8%、95.2%和96.4%，特异性分别为93.6%、94.0%、96.0%和97.6%，准确率分别为94.0%、93.4%、95.6%和97.0%。虽然人工智能模型诊断脑积水的敏感性、特异性、准确率明显低于副主任医师（P<0.05），但是与住院医师、主治医师均无统计学差异（P<0.05）。结论 本研究开发的人工智能模型可有效地识别脑积水的CT影像特点，具有较高的准确率。
|
| 关 键 词： | 脑积水 CT影像 诊断 人工智能 迁移学习算法 |
Application of artificial intelligence model established based on transfer learning algorithm in diagnosis of hydrocephalus |
| |
| CHEN Yu-hang,LIN Zhong-song,LI Xin-yu,ZHANG Liang,LI Xiao-pan,HAO Xiao-wei,CHENG Xiao-bing,ZHANG Hong-ri. ..Application of artificial intelligence model established based on transfer learning algorithm in diagnosis of hydrocephalus[J].Chinese Journal of Clinical Neurosurgery,2021,26(5):349-351. |
| |
| Authors: | CHEN Yu-hang LIN Zhong-song LI Xin-yu ZHANG Liang LI Xiao-pan HAO Xiao-wei CHENG Xiao-bing ZHANG Hong-ri |
| |
| Affiliation: | Department of Neurosurgery, First Affiliated Hospital, Henan University of Science and Technology, Luoyang 471000, China; 2.Shanghai Nanoperception Technology Co. LTD, Shanghai 201203, China; 3. Department of Radiology, First Affiliated Hospital, Henan Un |
| |
| Abstract: | Objective To explore the value of an artificial intelligence (AI) model established based on the transfer learning algorithm in the diagnosis of hydrocephalus. Methods The DICOM data of brain CT images of 1 250 normal adults and 1 250 patients with hydrocephalus were collected, and then were randomly divided into training set, verification set and test set at the ratio of 6:2:2. The data were preprocessed and the characteristic variables were labeled using Python tool, and then the AI model, which was established by a self-developed pixel-by-pixel ForrestNet-CNN algorithm, was used to extract and study the characteristic variables. The diagnostic results of AI model were compared with the results of imaging resident, attending physician and deputy chief physician, respectively. Results The sensitivities of the AI model, resident physician, attending physician, and deputy chief physician in diagnosing hydrocephalus were 94.4%, 92.8%, 95.2%, and 96.4%; the specificities were 93.6%, 94.0%, 96.0%, and 97.6%; the accuracy rates were 94.0%, 93.4%, 95.6% and 97.0%, respectively. The sensitivity, specificity and accuracy rate in diagnosing hydrocephalus of the AI model were significantly lower than those of the deputy chief physician (P<0.05). There was no statistical differences in the sensitivity, specificity and accuracy rate in diagnosing hydrocephalus among the AI model, the resident physician and the attending physician (P>0.05). Conclusions The AI model developed in this study can effectively identify the CT image characteristics of hydrocephalus with high accuracy. |
| |
| Keywords: | Hydrocephalus Artificial intelligence Transfer learning Diagnosis |
|
| 点击此处可从《中国临床神经外科杂志》浏览原始摘要信息 |
|
点击此处可从《中国临床神经外科杂志》下载全文 |
|
