[Application of convolutional neural network to risk evaluation of positive circumferential resection margin of rectal cancer by magnetic resonance imaging].

Objective: To explore the feasibility of using faster regional convolutional neural network (Faster R-CNN) to evaluate the status of circumferential resection margin (CRM) of rectal cancer in the magnetic resonance imaging (MRI). Methods: This study was registered in the Chinese Clinical Trial Registry (ChiCTR-1800017410). Case inclusion criteria: (1) the positive area of CRM was located between the plane of the levator ani, anal canal and peritoneal reflection; (2) rectal malignancy was confirmed by electronic colonoscopy and histopathological examination; (3) positive CRM was confirmed by postoperative pathology or preoperative high-resolution MRI. Exclusion criteria: patients after neoadjuvant therapy, recurrent cancer after surgery, poor quality images, giant tumor with extensive necrosis and tissue degeneration, and rectal tissue construction changes in previous pelvic surgery. According to the above criteria, MRI plain scan images of 350 patients with rectal cancer and positive CRM in The Affiliated Hospital of Qingdao University from July 2016 to June 2019 were collected. The patients were classified by gender and tumor position, and randomly assigned to the training group (300 cases) and the validation group (50 cases) at a ratio of 6:1 by computer random number method. The CRM positive region was identified on the T2WI image using the LabelImg software. The identified training group images were used to iteratively train and optimize parameters of the Faster R-CNN model until the network converged to obtain the best deep learning model. The test set data were used to evaluate the recognition performance of the artificial intelligence platform. The selected indicators included accuracy, sensitivity, positive predictive value, receiver operating characteristic (ROC) curves, areas under the ROC curves (AUC), and the time taken to identify a single image. Results: The accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of the CRM status determined by the trained Faster R-CNN artificial intelligence approach were 0.884, 0.857, 0.898, 0.807, and 0.926, respectively; the AUC was 0.934 (95% CI: 91.3% to 95.4%). The Faster R-CNN model's automatic recognition time for a single image was 0.2 s. Conclusion: The artificial intelligence model based on Faster R-CNN for the identification and segmentation of CRM-positive MRI images of rectal cancer is established, which can complete the risk assessment of CRM-positive areas caused by in-situ tumor invasion and has the application value of preliminary screening.目的： 探讨利用更快的区域卷积神经网络算法（Faster R-CNN）分析直肠癌磁共振（MRI）平扫图像以评估环周切缘（CRM）状态的可行性。 方法： 病例纳入标准：（1）CRM阳性区域位于提肛肌肛管平面至腹膜反折之间，电子结肠镜及组织病理学检查证实为直肠恶性肿瘤；（2）术后病理或术前高分辨MRI检查证实有CRM阳性。排除标准：新辅助治疗后、手术后复发、图像质量差伪影、肿瘤巨大及广泛坏死等组织变性以及既往盆腔手术史直肠临近组织结构改变的病例。根据以上标准，收集青岛大学附属医院2016年7月至2019年6月期间，350例CRM阳性直肠癌患者的MR平扫图像。将患者按照性别和肿瘤部位分类，应用计算机随机数字法以6∶1比例随机分配为训练组（300例）和测试组（50例）。利用LabelImg软件在T2加权成像（T2WI）图像上对CRM阳性区域进行标识。应用标识的训练组图像对Faster R-CNN模型进行迭代训练、优化参数，直到网络收敛获得最佳的深度学习模型。利用测试集数据评估人工智能平台的识别性能，选取指标包括准确度、敏感性、阳性预测值、受试者工作特征曲线和曲线下面积（AUC），以及识别单张图像所花费的时间。 结果： 训练后的Faster R-CNN模型判定CRM状态的准确度、敏感性、特异性、阳性预测值和阴性预测值分别为0.884、0.857、0.898、0.807和0.926；AUC为0.934（95%CI：91.3%~95.4%）。Faster R-CNN模型对单张图像的自动识别时间为0.2 s。 结论： 基于Faster R-CNN建立的对直肠癌CRM阳性MRI图像识别和分割的人工智能模型，可完成对原位肿瘤外侵引起CRM阳性区域的风险评估，具有初步筛选的应用价值。.