欢迎访问《中国医学影像技术》网站。

设为首页 | 加入收藏 | 联系我们

2025-06-15 星期日
首页 | 本刊简介 | 编委会 | 收录情况 | 投稿须知 | 期刊订阅 | 科学家精神 | 科学数据银行 | 广告招商

王府进,孟名柱,王欣,魏宁宁.改良YOLO-V5模型用于识别CT肠道造影所示炎性肠病[J].中国医学影像技术,2024,40(10):1593~1598

改良YOLO-V5模型用于识别CT肠道造影所示炎性肠病

Modified YOLO-V5 model for identifying inflammatory bowel disease on CT enterography

投稿时间：2024-02-29 修订日期：2024-05-14

DOI：10.13929/j.issn.1003-3289.2024.10.028

中文关键词: 肠道疾病诊断人工智能

英文关键词:intestinal diseases diagnosis artificial intelligence

基金项目:

作者	单位	E-mail
王府进	南京大学医学院附属盐城第一医院影像科, 江苏盐城 224000
孟名柱	南京医科大学附属常州第二人民医院影像科, 江苏常州 213164
王欣	南京大学医学院附属盐城第一医院影像科, 江苏盐城 224000
魏宁宁	南京医科大学附属常州第二人民医院影像科, 江苏常州 213164	19851971610@163.com

摘要点击次数: 369

全文下载次数: 143

中文摘要:

目的观察改良YOLO-V5模型识别CT肠道造影(CTE)中的炎性肠病(IBD)的价值。方法回顾性收集192例IBD患者及103例临床疑诊IBD而CTE未见异常者(无异常亚组)作为研究组,另以5例CD及3例UC为测试组。将含有表现为肠壁增厚的病变肠管或肠管无异常的CTE图像作为数据集(n=3 511)。按9 : 1比例将研究组CTE图像分为训练集(n=3 160,含CD亚组1 063幅、UC亚组931幅及无异常亚组1 166幅)与验证集(n=351,含CD亚组118幅、UC亚组103幅及无异常亚组130幅);以测试组25幅图像(CD 17幅、UC 8幅)为测试集。分类标注CTE中的CD病变肠管、UC病变肠管及无异常肠管;以改良YOLO-V5构建及训练包括YOLO-V5n、YOLO-V5s、YOLO-V5m、YOLO-V5l和YOLO-V5x在内的5个子模型,于测试集进行验证,并以精确度(Pr)、召回率(Rc)及平均精确度均值(mAP)评估其识别CTE中IBD病变肠管的效能。结果上述5个子模型的复杂程度逐次递增;YOLO-V5l及YOLO-V5x诊断效能较佳,前者识别训练集和验证集IBD病变肠管的整体Pr、Rc、mAP_{_0.5}及mAP_{_0.5:0.95}分别为0.97、0.93、0.96及0.91,后者分别为0.97、0.95、0.96及0.92。测试集中,YOLO-V5n子模型识别IBD病变肠管的效能较低,其识别CD与UC的mAP_{_0.5:0.95}分别为0.66与0.82,而YOLO-V5x子模型识别CD的mAP_{_0.5:0.95}达0.92、YOLO-V5l子模型识别UC的mAP_{_0.5:0.95}达0.91。结论基于改良YOLO-V5子模型YOLO-V5l和YOLO-V5x能有效识别CTE中的IBD病变肠管。

英文摘要:

Objective To investigate the value of modified YOLO-V5 model for identifying inflammatory bowel disease (IBD) displayed on CT enterography (CTE). Methods Totally 192 patients with IBD (103 cases of Crohn disease[CD subgroup] and 89 cases of ulcerative colitis[UC subgroup]) and 103 patients with clinically suspected IBD but CTE showed no abnormality (no abnormality subgroup) were retrospectively collected as study group, while 5 patients with CD and 3 with UC were collected as test group. CTE images with diseased intestinal tubes present as thickened intestinal wall or no abnormality intestinal tubes were selected as data set (n=3 511). CTE in study group were divided into training set (n=3 160, including 1 063 from CD subgroup, 931 from UC subgroup and 1 166 from no abnormality subgroup) and verification set (n=351, including 118 from CD subgroup, 103 from UC subgroup and 130 from no abnormality subgroup) at the ratio of 9∶1, while 25 CET images (17 from 5 cases of CD and 8 from 3 cases of UC) in test group were used as test set. Diseased tubes of CD, UC and no abnormality tubes were labeled. Then 5 sub-models, including YOLO-V5n, YOLO-V5s, YOLO-V5m, YOLO-V5l and YOLO-V5x were constructed and trained with modified YOLO-V5, and their efficacy were verified in test set. Precision (Pr), recall (Rc) and mean average precision (mAP) were used to evaluate the efficacy of each sub-model for identifying IBD lesions displayed on CTE. Results The complexity of the above 5 sub-models increased successively. YOLO-V5l and YOLO-V5x sub-model had better diagnostic efficacy, the overall Pr, Rc, mAP_{_0.5} and mAP_{_0.5:0.95} of the former for identifying IBD lesions in training and validation sets was 0.97, 0.93, 0.96 and 0.91, while of the latter was 0.97, 0.95, 0.96 and 0.92, respectively. In test set, the efficacy of YOLO-V5n sub-model for identifying IBD lesions was low, with mAP_{_0.5:0.95} of 0.66 and AUC of 0.82, whereas mAP_{_0.5:0.95} of YOLO-V5x sub-model for identifying CD was as high as 0.92, and of YOLO-V5l sub-model for identifying UC was as high as 0.91. Conclusion YOLO-V5l and YOLO-V5x sub-models based on modified YOLO-V5 could effectively identify IBD lesions displayed on CTE.

查看全文查看/发表评论下载PDF阅读器