深度学习和磁共振黑血血栓成像可用于下肢深静脉血栓分期预测  

Deep learning and black-blood magnetic resonance thrombus imaging can be used for predicting the staging of deep vein thrombosis in the lower limbs

作　　者：段丽芬 叶裕丰 陈秋梅 郭广源 黄益 DUAN Lifen;YE Yufeng;CHEN Qiumei;GUO Guangyuan;HUANG Yi(Department of Radiology,The Affiliated Panyu Central Hospital of Guangzhou Medical University,Guangzhou 511400,China;Department of Ultrasound,Guangzhou Panyu District Maternal and Child Health Hospital,Guangzhou 511499,China;Department of Minimally Invasive Intervention,The Affiliated Panyu Central Hospital of Guangzhou Medical University,Guangzhou 511400,China)

机构地区：[1]广州医科大学附属番禺中心医院放射科,广东广州511400 [2]广州市番禺区妇幼保健院超声科,广东广州511499 [3]广州医科大学附属番禺中心医院微创介入科,广东广州511400

出　　处：《分子影像学杂志》2024年第12期1335-1340,共6页Journal of Molecular Imaging

基　　金：广东省医学科研基金项目(A2024625);广州市科技计划项目(202201011638);广州市番禺区科技计划项目(2022-Z04-006)。

摘　　要：目的基于深度学习和磁共振黑血血栓成像(BTI)构建下肢深静脉血栓(DVT)的分期预测模型,并探讨其分期预测性能。方法回顾性收集2015年11月~2022年7月在广州市番禺区中心医院检查的196例患者的检查信息和BTI图像信息,随机性将数据集划分为3个部分:训练集占约70%(n=136),验证集与测试集各约占15%(n=30)。手动对实验组图像进行人工勾画矩形框,然后将对应的病变区域最小外接矩形框进行裁剪、统一尺寸、切片并输入深度学习模型中,基于ResNet50、ViT和EfficientNet构建3个下肢DVT分期预测模型,计算准确率、曲线下面积评估其预测性能。结果ResNet50、ViT和EfficientNet-b0在测试集上的准确度分别为0.693、0.733、0.787,EfficientNet-b0在测试集上展现出了最优的分类性能;在急性期、亚急性期及慢性期的曲线下面积分别为0.700(0.568~0.811)、0.778(0.652~0.875)、0.850(0.737~0.914)。结论结合磁共振黑血血栓成像图像,利用深度学习预测模型在DVT分期预测中具有一定的应用价值,这为DVT的精准分期提供了一个有效的技术路径。Objective To construct a staging prediction model on deep vein thrombosis(DVT)based on the deep learning and black-blood magnetic resonance thrombus imaging(BTI),and investigate its prediction value.Methods A retrospective observational study was conducted,where clinical data and BTI from 196 patients admitted to Guangzhou Panyu Central Hospital from November 2015 to July 2022 were collected and analyzed.The dataset was split into a training set(70%,n=136),a validation set(15%,n=30),and a test set(15%,n=30).The experimental group were annotated in rectangular boxes manually,then the corresponding minimum bounding rectangular boxes of the lesion areas were cropped,resized,and sliced,and input to the deep learning model.The three models,ResNet50,Vit and EfficientNet,were established for lower limb staging prediction.Their predict value were compared by accuracy rate and the area under the curve(AUC).Results The accuracy of ResNet50,Vit and EfficientNet-b0 in the testing set were 0.693,0.733,0.787.The EfficientNet-b0 outperforms than other two models in the test set.The area under the curve of the acute,sub-acute and chronic phase were 0.700(0.568-0.811),0.778(0.652-0.875),0.850(0.737-0.914),respectively.Conclusion Deep learning combined with BTI has certain application values in staging prediction of DVT.It provides an effective technique for the precisive staging for DVT.

关 键 词：深度学习 磁共振成像 下肢深静脉血栓 血栓分期 

分 类 号：R543.6[医药卫生—心血管疾病] R445.2[医药卫生—内科学]