目的:克罗恩病(CD)的定性发现对于可靠地报告和量化可能具有挑战性。我们评估了机器学习方法,以标准化回肠CD的常见定性发现的检测,并确定在CT小肠造影(CTE)上发现空间定位。
方法:纳入2016年至2021年单中心回顾性研究的回肠CD和CTE患者。165CTE由两名受过研究金训练的腹部放射科医师审查了五个定性CD发现的存在和空间分布:壁画增强,壁画分层,狭窄,壁厚,和肠系膜脂肪绞合。开发了使用自动提取的专家指导的肠特征和无偏卷积神经网络(CNN)的随机森林(RF)集成模型来预测定性发现的存在。使用曲线下面积(AUC)评估模型性能,灵敏度,特异性,准确度,和kappa协议统计。
结果:在165名受试者中,进行了29,895项个人定性发现评估,放射科医生对定位的一致性从好到非常好(κ=0.66到0.73),除了肠系膜脂肪绞合(κ=0.47)。射频预测模型具有优异的性能,总体AUC,灵敏度,特异性分别为0.91、0.81和0.85。用于CD发现定位的RF模型和放射科医师协议近似放射科医师之间的协议(κ=0.67至0.76)。没有疾病知识的无偏CNN模型与使用专家定义的成像特征的RF模型具有非常相似的性能。
结论:用于CTE图像分析的机器学习技术可以识别存在,location,以及与经验丰富的放射科医生性能相似的定性CD发现的分布。
OBJECTIVE: Qualitative findings in Crohn\'s disease (CD) can be challenging to reliably report and quantify. We evaluated machine learning methodologies to both standardize the detection of common qualitative findings of ileal CD and determine finding spatial localization on CT enterography (CTE).
METHODS: Subjects with ileal CD and a CTE from a single center retrospective study between 2016 and 2021 were included. 165 CTEs were reviewed by two fellowship-trained abdominal radiologists for the presence and spatial distribution of five qualitative CD findings: mural enhancement, mural stratification, stenosis, wall thickening, and mesenteric fat stranding. A Random Forest (RF) ensemble model using automatically extracted specialist-directed bowel features and an unbiased convolutional neural network (CNN) were developed to predict the presence of qualitative findings. Model performance was assessed using area under the curve (AUC), sensitivity, specificity, accuracy, and kappa agreement statistics.
RESULTS: In 165 subjects with 29,895 individual qualitative finding assessments, agreement between radiologists for localization was good to very good (κ = 0.66 to 0.73), except for mesenteric fat stranding (κ = 0.47). RF prediction models had excellent performance, with an overall AUC, sensitivity, specificity of 0.91, 0.81 and 0.85, respectively. RF model and radiologist agreement for localization of CD findings approximated agreement between radiologists (κ = 0.67 to 0.76). Unbiased CNN models without benefit of disease knowledge had very similar performance to RF models which used specialist-defined imaging features.
CONCLUSIONS: Machine learning techniques for CTE image analysis can identify the presence, location, and distribution of qualitative CD findings with similar performance to experienced radiologists.