引用文本：陈小佩, 徐凡, 李艳梅, 等. 基于术前增强计算机断层扫描的深度学习模型预测胃癌人表皮生长因子受体2表达状态的研究[J/CD]. 消化肿瘤杂志（电子版）, 2026, 18(1): 89-95.

作者：陈小佩，徐凡，李艳梅，叶郭锡，廖智，张创嘉，步军

单位：广州市红十字会医院放射科，广东 广州 510220

Authors：Chen Xiaopei, Xu Fan, Li Yanmei, Ye Guoxi, Liao Zhi, Zhang Chuangjia, Bu Jun

Unit：Department of Radiology, Guangzhou Red Cross Hospital, Guangzhou 510220, Guangdong, China

摘要：

目的 探讨基于增强计算机断层扫描（computed tomography, CT）的深度学习模型在术前预测胃癌患者人表皮生长因子受体2（human epidermal growth factor receptor 2, HER2）表达状态的临床价值。方法 本研究为回顾性研究，选取2015年1月1日至2023年8月31日在广州市红十字会医院接受术前增强CT检查并于胃切除术后检测HER2状态的265例胃癌患者，采用随机数字表法按照约6∶2∶2的比例分为训练集165例、验证集54例、测试集46例。提取患者术前增强CT动脉期（arterial phase, AP）、静脉期（venous phase, VP）、延迟期（delayed phase, DP）影像学特征，分别构建单期相及多期相融合（multi-phase fusion, MP）深度学习模型，通过受试者操作特征曲线（receiver operating characteristic curve, ROC曲线）评估各模型预测胃癌 HER2 表达状态的性能。结果 在训练集、验证集、测试集中，HER2阳性病例占比分别为18.8%（31/165）、24.1%（13/54）、19.6%（9/46），阴性病例占比分别为81.2%（134/165）、75.9%（41/54）、80.4%（37/46）。在验证集中，AP、VP、DP、MP模型的曲线下面积（area under the curve, AUC）依次为0.670（95%CI 0.666~0.673）、0.875（95%CI 0.871~0.877）、0.636（95%CI 0.635~0.641）、0.895（95%CI 0.894~0.896）；在测试集中，上述4个模型的AUC分别为0.738（95%CI 0.734~0.740）、0.671（95%CI 0.666~0.673）、0.636（95%CI 0.635~0.641）、0.783（95%CI 0.782~0.786），MP模型在验证集（与AP、VP、DP模型比较，Z=121.2、12.5、160.6，均P＜0.001）和测试集（与AP、VP、DP模型比较，Z=24.5、54.5、79.9，均P＜0.001）中的AUC均最高。结论 基于增强CT的深度学习模型可有效在术前预测胃癌HER2表达状态，为临床术前评估和治疗决策提供可靠的辅助方法。

关键词：胃癌；人表皮生长因子受体2；深度学习；计算机断层扫描

Abstract：

Objective  To explore the clinical value of a deep learning (DL) model based on contrast-enhanced computed tomography (CECT) in the preoperative prediction of human epidermal growth factor receptor 2 (HER2) expression status in patients with gastric cancer (GC). Method This retrospective study included 265 GC patients who underwent CECT preoperatively and HER2 expression detection after gastrectomy from Guangzhou Red Cross Hospital between January 1, 2015, and August 31, 2023. The patients were divided into a training set (165 cases), a validation set (54 cases), and a test set (46 cases) by random number table at a ratio of 6∶2∶2 approximately. Imaging features were extracted from the arterial phase (AP), venous phase (VP), and delayed phase (DP) of preoperative CECT images. Single-phase and multi-phase fusion (MP) DL models were constructed respectively, and the performance of each model in predicting HER2 expression in GC was evaluated and compared using receiver operating characteristic (ROC) curves. Result In the training, validation, and test sets, the proportions of HER2-positive cases were 18.8% (31/165), 24.1% (13/54), and 19.6% (9/46), respectively, while the proportions of HER2-negative cases were 81.2% (134/165), 75.9% (41/54), and 80.4% (37/46), respectively. In the validation set, the area under the curve (AUC) of the AP, VP, DP, and MP models were 0.670 (95%CI 0.666-0.673), 0.875 (95%CI 0.871-0.877), 0.636 (95%CI 0.635-0.641), and 0.895 (95%CI 0.894-0.896), respectively. In the test set, the AUC of these four models were 0.738 (95%CI 0.734-0.740), 0.671 (95%CI 0.666-0.673), 0.636 (95%CI 0.635-0.641), and 0.783 (95%CI 0.782-0.786), respectively. Notably, the MP model achieved the highest AUC in both the validation set (compared with AP, VP, and DP model, Z=121.2, 12.5, 160.6, respectively, all P＜0.001) and the test set (compared with AP, VP, and DP model, Z=24.5, 54.5, 79.9, respectively, all P＜0.001). Conclusion The DL model based on CECT can effectively preoperatively predict HER2 expression status in GC, thereby providing a reliable auxiliary method for clinical preoperative evaluation and treatment decision-making.

Key words：Gastric cancer; Human epidermal growth factor receptor 2; Deep learning; Computed tomography

注：网络优先发布