|
| 不同机器学习模型鉴别结肠型克罗恩病与溃疡性结肠炎的价值 |
| The value of different machine learning models in discriminating colonic Crohn’s disease from ulcerative colitis |
| 投稿时间：2022-05-12 |
| DOI：10.3969/j.issn.1000-0399.2023.01.004 |
| 中文关键词: 结肠型克罗恩病 溃疡性结肠炎 影像组学 机器学习模型 |
| 英文关键词: Colonic Crohn’s disease Ulcerative colitis Radiomics Machine learning models |
| 基金项目:安徽省学术技术带头人科研项目(编号:2021D299) |
|
| 摘要点击次数: 1088 |
| 全文下载次数: 306 |
| 中文摘要: |
| 目的 比较不同的机器学习模型在鉴别结肠型克罗恩病(CCD)与溃疡性结肠炎(UC)中的价值。方法 收集2019年6月至2021年12月在安徽医科大学第一附属医院消化科就诊的有完整CT小肠成像(CTE)且经病理证实炎症性肠病(IBD)患者44例(CCD 25例,UC 19例)。利用ITK-Snap软件在静脉期病灶最明显肠段进行勾画,共计勾画106个病变肠段(CCD 58个、UC 48个)。利用AK软件提取勾画区影像组学特征,以7∶[KG-*2]3比例随机分为训练集和测试集;对训练集用Correlation_xx和MultiVariate_Logistic算法进行数据降维,筛选组间差异明显的影像组学特征构建6种机器学习模型,用测试集的特征对其进行验证。结果 175种组学特征中有4种组间差异有统计学意义(P＜0.05)。6种模型中有4种模型曲线下面积均＞0.90。训练集中邻近算法(KNN)模型鉴别CCD与UC的受试者工作特征曲线下面积(AUC)为0.958(95% CI:0.917～0.992),准确率、特异度、灵敏度分别为87.7%、100%和72.7%;在测试集的AUC为0.904(95% CI:0.792～0.996),准确率、特异度和灵敏度分别为87.9%、88.9%和86.7%。结论 4种常用的机器学习模型在鉴别CCD与UC中均有良好的表现;其中KNN模型稳定性好,准确性更高。 |
| 英文摘要: |
| Objective To compare the value of different machine learning models in distinguishing colonic Crohn’s disease (CCD) from ulcerative colitis (UC). Methods A total of 44 patients (25 CCD and 19 UC) with complete CT enterography (CTE) imaging and pathologically confirmed inflammation bowel diseases (IBD) were collected. Itk-snap software was used to outline the intestinal segments with the most obvious lesions at the venous phase imaging, and a total of 106 intestinal segments of lesions were delineated (58 CCD and 48 UC). The radiomics features of the delineation area were extracted by A.K. software and randomly divided into training set and test set in a ratio of 7∶[KG-*2]3. The Correlation_xx and MultiVariate_Logistic algorithms were used to reduce the dimensionality of the training set, and six machine learning models were constructed by screening the radiomics features with obvious differences between the groups, and then verified by the features of the test set. Results Four of the 175 radiomics characteristics differed significantly between groups. Four of the six models have areas under the curve>0.90. The Area Under Curve (AUC) (95%CI), accuracy, specificity, and sensitivity of KNN model in distinguishing CCD from UC in training set were 0.958 (0.917~0.992), 87.7%, 100% and 72.7%, respectively. In the test set, AUC (95%CI) was 0.904(0.792~0.996), accuracy, specificity and sensitivity were 87.9%, 88.9% and 86.7%, respectively.Conclusion The machine learning models have good performance in discriminating CCD and UC. KNN model has better stability and accuracy. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
