Background

Imaging is essential in the assessment of Crohn’s disease (CD) severity, phenotype, and therapeutic response. However, qualitative findings can be limited by interobserver variation and ambiguous feature definitions. Our aim was to evaluate computational approaches for identifying qualitative disease features using CT-enterography (CTE).

Methods

CD subjects with ileal CD and CTE imaging between 2009 and 2017 were retrospective identified at a single tertiary care centre. CTE scans were reviewed by two fellowship-trained abdominal radiologists who labelled diseased and normal bowel transitions agreeing on definitions of qualitative findings prior to labelling. Computed intestinal features were used by machine learning methods (k-nearest neighbour, support vector machines, random forest) to model regions of diseased bowel and predict qualitative findings with 5-fold cross validation. Cohen’s kappa with quadratic weighting was used to assess agreement between radiologists and model predictions.

Results

In 206 unique patients, 548 small bowel segments underwent paired radiologist review for qualitative imaging findings. Automated localisation of diseased vs. normal bowel segments had excellent performance, with an AUC, PPV, and NPV of 0.922, 0.924, and 0.918, respectively (Figure 1). Radiologist-to-Radiologist and Radiologist-to-automated prediction agreement on qualitative findings are shown in Table 1. Agreement on the degree of mural enhancement between radiologists was very good (k = 0.75,95% CI 0.68–0.82), with nearly identical agreement (k = 0.75, 95% CI 0.72–0.79) between radiologists and automated grading models.Table 1.Agreement of qualitative imaging findings of Crohn’s disease between radiologists and automated methodsRadiologist–radiologistAutomated prediction performanceKappa (95% CI)Kappa (95% CI)PPVNPVAUCBowel Wall Thickening0.75 (0.69–0.81)0.77 (0.74–0.80)85.6%91.1%0.922Lumen Narrowing0.55 (0.49–0.61)0.65 (0.60–0.69)73.9%90.2%0.859Mural Stratification0.76 (0.71–0.81)0.51 (0.46–0.56)65.2%87.0%0.809Fat Stranding0.34 (0.27–0.41)0.26 (0.18–0.33)77.4%88.7%0.879Vascular Engorgement0.61 (0.51–0.71)0.14 (0.06–0.22)47.4%91.6%0.909Figure 1.Example figure of a section of distal ileum affected by Crohn’s disease with radiologist and predicted qualitative findings shown. Automated disease segment identification was very good with an AUC of 0.922, reflecting both disease/non-disease prediction and spatial localisation of disease.

Conclusion

Computer vision methods have excellent performance for automatically distinguishing diseased from normal ileum and show potential for qualitative disease assessments of Crohn’s disease on CTEs.

中文翻译：

---

P158使用机器学习在CT肠镜上半自动检测克罗恩病活动的定性特征

背景

成像对于评估克罗恩病（CD）的严重程度，表型和治疗反应至关重要。但是，定性结果可能会受到观察者之间的差异和模棱两可的特征定义的限制。我们的目标是评估使用CT肠胃造影（CTE）识别定性疾病特征的计算方法。

方法

在单个三级护理中心回顾性鉴定了2009年至2017年间具有回肠CD和CTE影像学的CD受试者。两名接受过研究金培训的腹部放射科医生对CTE扫描进行了审查，他们对患病的和正常的肠转移进行了标记，并在标记之前就定性发现的定义达成了一致。通过机器学习方法（k近邻，支持向量机，随机森林），使用计算出的肠道特征对患病肠区域进行建模，并通过5倍交叉验证来预测定性结果。二次加权的Cohen卡伯（Kappa）用于评估放射科医生与模型预测之间的一致性。

结果

在206位独特的患者中，对548个小肠段进行了放射放射科医生复查，以获取定性影像学检查结果。患病肠段与正常肠段的自动定位具有出色的性能，AUC，PPV和NPV分别为0.922、0.924和0.918（图1）。表1显示了放射科医生对放射科医生和放射科医生对定性发现的自动预测协议。放射科医生之间壁画增强程度的协议非常好（k = 0.75,95％CI 0.68–0.82），几乎相同放射科医生与自动分级模型之间的一致性（k = 0.75，95％CI 0.72–0.79）。表1.放射科医生与自动化方法之间克罗恩病定性影像学发现的一致性放射科医生-放射科医生自动化预测性能卡帕（95％CI）卡帕（95％CI PPVNPVAUC牛墙增厚0。75（0.69–0.81）0.77（0.74–0.80）85.6％91.1％0.922流明变窄0.55（0.49–0.61）0.65（0.60–0.69）73.9％90.2％0.859壁垒0.76（0.71-0.81）0.51（0.46– 0.56）65.2％87.0％0.809脂肪搁浅0.34（0.27–0.41）0.26（0.18–0.33）77.4％88.7％0.879血管充盈0.61（0.51-0.71）0.14（0.06-0.22）47.4％91.6％0.909图1。放射科医师对受克罗恩病影响的回肠远端切片的示例图，并显示了预测的定性结果。疾病区段的自动识别非常好，AUC为0.922，既反映了疾病/非疾病预测，又反映了疾病的空间定位。51-0.71）0.14（0.06-0.22）47.4％91.6％0.909图1.放射科医师对受克罗恩病影响的回肠远端切片的示例图，并显示了预期的定性结果。疾病区段的自动识别非常好，AUC为0.922，既反映了疾病/非疾病预测，又反映了疾病的空间定位。51-0.71）0.14（0.06-0.22）47.4％91.6％0.909图1.放射科医师对受克罗恩病影响的回肠远端切片的示例图，并显示了预期的定性结果。疾病区段的自动识别非常好，AUC为0.922，既反映了疾病/非疾病预测，又反映了疾病的空间定位。

结论

计算机视觉方法具有出色的性能，可自动区分疾病与正常回肠，并显示出在CTE上克罗恩病定性疾病评估的潜力。