Li Hui, Mo Yan, Huang Chencui, Ren Qingguo, Xia Xiaona, Nan Xiaomin, Shuai Xinyan, Meng Xiangshui
Department of Radiology, Qilu Hospital (Qingdao), Cheeloo College of Medicine, Shandong University, Qingdao, China.
Deepwise AI Lab, Beijing Deepwise & League of PHD Technology Co., Ltd., Beijing, China.
Ann Transl Med. 2021 Apr;9(7):572. doi: 10.21037/atm-21-1023.
We established and evaluated a radiomics nomogram based on multislice computed tomography (MSCT) arterial phase contrast-enhanced images to distinguish between Crohn's disease (CD) and ulcerative colitis (UC) objectively, quantitatively, and reproducibly.
MSCT arterial phase-enhancement images of 165 lesions (99 CD, 66 UC) in 87 patients with inflammatory bowel disease (IBD) confirmed by endoscopy or surgical pathology were retrospectively analyzed. A total of 132 lesions (80%) were selected as the training cohort and 33 lesions (20%) as the test cohort. A total of 1648 radiomic features were extracted from each region of interest (ROI), and the Pearson correlation coefficient and tree-based method were used for feature selection. Five machine learning classifiers, including logistic regression (LR), support vector machine (SVM), random forest (RF), stochastic gradient descent (SGD), and linear discriminative analysis (LDA), were trained. The best classifier was evaluated and obtained, and the results were transformed into the Rscore. Three clinical factors were screened out from 8 factors by univariate analysis. The logistic regression method was used to synthesize the significant clinical factors and the Rscore to generate the nomogram, which was compared with the clinical model and LR model.
Among all machine learning classifiers, LR performed the best (AUC =0.8077, accuracy =0.697, sensitivity =0.8, specificity =0.5385), SGD model had the second best performance (AUC =0.8, accuracy =0.6667, sensitivity =0.75, specificity =0.5385), and the DeLong test results showed that there was no significant difference between LR and SGD (P=0.465>0.05), while the other models performed poorly. Texture features had the greatest impact on classification results among all imaging features. The significant features of the LR model were used to calculate the Rscore. The 3 significant clinical factors were perienteric edema or inflammation, CT value of arterial phase-enhancement (AP-CT value), and lesion location. Finally, a nomogram was constructed based on the 3 significant clinical factors and the Rscore, whose AUC (0.8846) was much higher than that of the clinical model (0.6154) and the LR model (0.8077).
The nomogram is expected to provide a new auxiliary tool for radiologists to quickly identify CD and UC.
我们建立并评估了一种基于多层螺旋计算机断层扫描（MSCT）动脉期增强图像的放射组学列线图，以客观、定量且可重复地鉴别克罗恩病（CD）和溃疡性结肠炎（UC）。
回顾性分析了87例经内镜或手术病理确诊的炎症性肠病（IBD）患者的165个病灶（99个CD，66个UC）的MSCT动脉期增强图像。总共132个病灶（80%）被选为训练队列，33个病灶（20%）被选为测试队列。从每个感兴趣区域（ROI）提取了总共1648个放射组学特征，并使用Pearson相关系数和基于树的方法进行特征选择。训练了包括逻辑回归（LR）、支持向量机（SVM）、随机森林（RF）、随机梯度下降（SGD）和线性判别分析（LDA）在内的五种机器学习分类器。对最佳分类器进行评估并获得结果，并将其转换为Rscore。通过单因素分析从8个因素中筛选出3个临床因素。使用逻辑回归方法将显著临床因素和Rscore进行综合以生成列线图，并将其与临床模型和LR模型进行比较。
在所有机器学习分类器中，LR表现最佳（AUC =0.8077，准确率 =0.697，敏感度 =0.8，特异度 =0.5385），SGD模型表现次之（AUC =0.8，准确率 =0.6667，敏感度 =0.75，特异度 =0.5385），DeLong检验结果显示LR和SGD之间无显著差异（P =0.465>0.05），而其他模型表现较差。在所有影像特征中，纹理特征对分类结果影响最大。使用LR模型的显著特征来计算Rscore。3个显著临床因素为肠周水肿或炎症、动脉期增强CT值（AP-CT值）和病灶位置。最后，基于3个显著临床因素和Rscore构建了列线图，其AUC（0.8846）远高于临床模型（0.6154）和LR模型（0.8077）。
该列线图有望为放射科医生快速鉴别CD和UC提供一种新的辅助工具。
