CT影像组学列线图预测胃癌内镜活检与术后病理学分型差异

目的 评估基于CT影像组学的列线图预测胃癌内镜活检与术后病理Lauren分型差异的价值。方法 回顾性分析126例经手术病理确诊的胃癌患者，根据内镜活检与术后病理结果是否一致将其分为一致组（n=77）与不一致组（n=49），同时按2∶1比例分为训练集与验证集。筛选临床预测因子，构建临床预测模型；于静脉期CT图像提取影像组学特征，采用L1正则化方法对特征进行筛选，以决策树、随机森林、逻辑回归3种机器学习（ML）算法构建影像组学模型；基于临床及最佳影像组学ML模型构建列线图；评估各模型及列线图的预测效能及其临床价值。结果 患者年龄、血小板计数、动脉期CT值是预测胃癌内镜活检与术后病理分型不一致的独立因子。3种ML模型中，随机森林预测效能较好，其在训练集与验证集中的曲线下面积（AUC）分别为0.835及0.724。临床模型、影像组学模型及列线图在训练集的AUC分别为0.764、0.835及0.884，在验证集分别为0.760、0.724及0.841；列线图在训练集与验证集均显示出较好的拟合度及临床实用性。结论 基于CT影像组学列线图用于预测胃癌内镜活检与术后病理Lauren分型不一致性具有潜在临床应用价值。

Nomogram based on CT radiomics for predicting pathological types of gastric cancer: Difference between endoscopic biopsy and postoperative pathology

Objective To observe the value of CT radiomics-based nomogram for predicting difference of Lauren types of gastric cancers between endoscopic biopsy and postoperative pathology. Methods Totally 126 patients with gastric cancer diagnosed by surgical pathology were retrospectively analyzed. The patients were divided into concordant group （n=77） and inconsistent group （n=49） according to the concordance between endoscopic biopsy and postoperative pathology results or not， also divided into training set and validation set at the ratio of 2∶1. Clinical predictors were screened， then a clinical prediction model was constructed. Radiomics features were extracted based on venous-phase CT images and screened using L1 regularization. Radiomics models were constructed using 3 machine learning （ML） algorithms， i.e. decision trees， random forests and logistic regression. The nomogram based on clinical and the best ML radiomics model was constructed， and the efficacy and clinical utility of the above models and nomogram for predicting inconsistency of Lauren types of gastric cancers between endoscopic biopsy and postoperative pathology were evaluated. Results Patients’ age， platelet count， and arterial-phase CT values of tumors were all independent predictors of inconsistency between endoscopic biopsy and postoperative pathology of Lauren types of gastric cancer. CT radiomics model using random forests algorithm showed better predictive efficacy among 3 ML models，with the area under the curve （AUC） of 0.835 in training set and 0.724 in validation set， respectively. The AUC of clinical model， radiomics model and the nomogram in training set was 0.764， 0.835 and 0.884， while was 0.760， 0.724 and 0.841 in validation set， respectively. In both training set and validation set， the nomogram showed a good fit and considerable clinical utility. Conclusion CT radiomics-based nomogram had potential clinical application value for predicting inconsistency of Lauren types of gastric cancers between endoscopic biopsy and postoperative pathology.