【摘要】 目的:探讨基于CT肠道成像(CTE)利用机器学习方法提取的影像特征在评估活动期溃疡性结肠炎(UC)病变程度中的价值。方法:将2017年9月-2020年9月在本院首诊为UC的157例患者患者纳入研究。根据Mayo临床评分方法,将所有患者分为轻度组(3～5分)40例、中度组(6～10分)52例、重度组(11～12分)65例。所有患者行肠镜及CTE检查。基于CTE图像记录每例患者13个影像学征象(病变范围、肠壁增厚、黏膜分层、肠壁异常强化、肠系膜充血、直肠周围脂肪沉淀、淋巴结增大、黏膜囊泡、肠腔狭窄、结肠袋消失、肠黏膜息肉、靶征和梳齿征)的出现情况。将所有患者按照5:5分为训练组和验证组,基于训练组数据,以Mayo分组为分类标签,对CTE征象进行特征筛选后纳入GBM分类器,利用机器学习方法构建CTE征象诊断模型,并利用验证组的数据对其进行验证。结果:经特征筛选后,将10个有统计学意义的CTE征象纳入GBM分类器,采用机器学习的方法构建CTE征象诊断模型,其在训练集中鉴别轻度与中重度、中度与重度、重度与轻中度的AUC分别为0.99、0.99和1.00,在验证组中相应AUC为0.99、0.96和0.98;在训练组合验证组中评估UC病变程度的总体诊断符合率分别为0.921(95%CI:0.8102～0.9553)和0.887(95%CI:0.7617～0.9274)。结论:利用多分类器机器学习的方法基于CTE影像特征构建的诊断模型可用于评估UC患者的病变程度。更多还原

【Abstract】 Objective:To explore the value of image features extracted by machine learning method based on CT intestinal imaging(CTE) in evaluating the severity of active ulcerative colitis(UC).Methods:The patients who were first diagnosed with UC from September 2017 to September 2020 in our hospital were enrolled in this study and all underwent colonoscopy and CTE examination.Each patients were assessed using Mayo Clinical Score.A total of 157 patients were enrolled in this study, including 40 patients with mild UC(Mayo score: 3～5),52 patients with moderate UC(Mayo score: 6～10),and 65 patients with severe UC(Mayo score: 11～12).Then for each patients, 13 features(extended range of UC,bowel wall thickening, mural straitification, mural hyperenhancement, mesentric hyperemia, periectal standing, lymph mode enlargement, mucosal buddles, luminal narrowing, loss of huastration, intestinal psendopolyp, target sign and comb sign) on CTE images were analyzed and recorded.All patients were divided into training group and verification group according to a ratio of 5:5.In the training group, patients were grouped by the extent of the UC based on Mayo score, and the difference of CTE features among the three groups were compared and significant features were selected out, then they were included into the GBM classifier for constructing CTE feature diagnostic model using machine learning method.The data of verification group were used for verification.Results:After feature selection, 10 features were finally selected out to construct CTE feature model using GBM classifier with machine learning.In the training set, the AUCs of the CTE feature label for differentiating mild from moderate and severe UC,moderate from severe UC,and severe from mild and moderate UC were 0.99,0.99 and 1.00,respectively; and in the test group, those were 0.99,0.96 and 0.98,respectively.The overall diagnostic accuracy of the diagnostic model in the training group and test group when assessing different degrees of UC was 0.921(95% confidence interval: 0.8102～0.9553) and 0.887(95%CI:0.7617～0.9274).Conclusion:This group of research uses the multi-classifier deep learning machine learning method for the first time to build a model based on the image features extracted by CTE to evaluate UC patients with different Mayo scores, in order to provide a clinically non-invasive diagnosis method for UC patients.更多还原