摘要
目的利用基于高清T2WI的深度学习模型来鉴别直肠癌直肠系膜内转移淋巴结和非转移淋巴结。方法收集2016年6月至2021年4月山东省千佛山医院就诊的直肠占位性病变患者166例,治疗前均接受3.0T磁共振扫描,包括高分辨T2WI(HR-T2WI)序列,并在2周内接受全直肠系膜切除术(TME)。术后由影像科和病理科医生共同对直肠系膜内的淋巴结进行定位和定性。将入组患者的淋巴结按照7∶3的比例随机分为训练组和测试组。提取入组患者的HR-T2WI图像,并对直肠系膜内淋巴结进行感兴趣区(ROI)标记。应用随机翻转和添加随机噪声两种数据增强方式来增加鲁棒性。建立包含5个卷积模块和2个全连接层的卷积神经网络(CNN)模型,并分别进行训练和测试。评价指标包括ROC曲线、AUC、准确性、敏感性和特异性。结果最终获得604个淋巴结(298个良性和306个恶性)。分别将训练集淋巴结(215恶性+205良性)和测试集淋巴结(91恶性+93良性)进行逆行深度学习CNN模型训练后,得到训练集和测试集的AUC分别为0.910、0.820。测试集准确率、敏感性和特异性分别为0.725、0.698和0.752。结论基于HR-T2WI序列深度学习方法可以用来鉴别直肠癌直肠系膜内淋巴结状态。
Objective To distinguish metastatic and non-metastatic mesenteric lymph nodes(LNs)in rectal cancer using deep learning based on high-resolution T2WI.MethodsA total of 166 patients with rectal space-occupying lesions were enrolled in Shandong Provincial Qianfoshan Hospital from June 2016 to April 2021.All patients underwent 3.0T magnetic resonance scanning, including high-resolution T2WI(HR-T2WI)sequence, and underwent total mesorectal excision(TME)within two weeks.The LNs within the mesentery were located and characterized by radiologist and pathologist.The LNs were randomly divided into training set and test set according to the ratio of 7∶3.HR-T2WI images were extracted from the enrolled patients, and the mesorectal LNs were labeled to get region of interest(ROI). Random flipping and adding random noise were used to increase the robustness.A convolutional neural network(CNN)model consisting of five convolution modules and two fully connected layers was established, and was trained and tested respectively.The evaluation indexes included ROC curve, AUC, accuracy, sensitivity and specificity.ResultsA total of 604 LNs(298 benign and 306 malignant)were finally obtained.After retrograde deep learning CNN model training of the LNs in the training set(215 malignant+ 205 benign)and the LNs in the test set(91 malignant+ 93 benign), the AUCs of the training set and the test set were 0.910 and 0.820 respectively.The accuracy, sensitivity and specificity of the test set were 0.725, 0.698 and 0.752 respectively.ConclusionThe deep learning method based on HR-T2WI sequence can be used to identify the status of mesorectal LNs in rectal cancer.
作者
贾玉萍
宋歌声
李爱银
Yuping Jia;Gesheng Song;Aiyin Li(Department of Radiology,First Affiliated Hospital of Shandong First Medical University&Shandong Provincial Qianfoshan Hospital,Jinan 250014,China)
出处
《中华消化病与影像杂志（电子版）》
2022年第6期342-347,共6页
Chinese Journal of Digestion and Medical Imageology(Electronic Edition)
基金
山东省科技发展计划(2014GSF118086)
济南市科技计划(201907034)
作者简介
通信作者:宋歌声,Email:songgesheng921@126.com
