摘要 目的探讨基于深度神经网络的目标检测技术在腹部双源CT胆囊癌辅助识别系统的临床应用价值。方法选取2017年1月至2019年12月上海交通大学医学院附属新华医院普外科、吉林大学第一医院肝胆胰外一科和吉林大学中日联谊医院普外科收治的88例病理学检查诊断明确的胆囊癌,28例慢性胆囊炎胆囊结石病人和29例正常胆囊(影像学检查胆囊正常)病人,均行腹部双源CT检查。随机选取101例作为训练组,29例作为验证组,15例作为测试组。首先,利用已标注的10409张腹部双源CT图像对Mask R-CNN模型进行学习,从而建立自动胆囊癌辅助识别系统。然后对验证组的2974张CT图像通过专业的医师对其进行判断识别,与Mask R-CNN得出的结果进行对比分析。通过不同交并比阈值(IoU)下的平均检测精度(AP)和平均召回率(AR)来对性能进行评估。结果计算机通过学习组不断迭代训练,Mask R-CNN的损失函数值收敛,诊断误差不断降低。在IoU为0.5时,Mask R-CNN的边界框和掩膜的AP分别为0.929和0.929,IoU为0.75时的边界框和掩膜AP分别为0.901和0.890,IoU为0.5:0.95时的边界框和掩膜AP分别为0.723和0.707,平均召回率分别为0.794和0.774,模型的性能良好。结论基于深度神经网络的Mask R-CNN胆囊癌辅助识别系统具有较高的准确率和性能,可辅助进行临床诊断。 Objective network Mask R-CNN.MethodsThe research selected 88 patients with gallbladder cancer,28 patients with chronic cholecystitis with gallbladder stones and 29 patients with normal gallbladder tissues who underwent abdominal dualsource CT examination and diagnosed by pathology from January 2017 to December 2019 in Department of General Surgery,Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine,First Department of Hepatobiliary and Pancreatic Surgery of Jilin University and Department of General Surgery,China-Japan Friendship Hospital of Jilin University.One hundred and onecases as the training group,29 cases as theverification group and 15 cases as the test groupwere selected randomly.Firstly,using deeplearning technique,researchers trained the 10409 abdominal dual-source CT images data oflearning group with convolution neural network tosimulate the judgment process of radiologists,andestablished an artificial intelligence automaticrecognition system for gallbladder cancer.Then,2974 images of the validation group wereclinically validated.The performance was evaluated by the average detection accuracy(AP)and average recall(AR)under different intersection ratio thresholds(Io U).ResultsAfter continuous iteration training of the learning group data,the loss function value of Mask R-CNNdecreased continuously,and the diagnostic error decreased continuously.When the Io U was 0.5,the AP of the boundingbox and mask of Mask R-CNN were 0.929 and 0.929,respectively.When the Io U was 0.75,the AP of the bounding boxand mask were 0.901 and 0.890,respectively.When the Io U was 0.5:0.95,the AP of the bounding box and mask were0.723 and 0.707,and the AR was 0.794 and 0.774;respectively.It suggested the performance of the model was good.ConclusionThe Mask R-CNN for gallbladder cancer automatic recognition system based on deep neural network hashigh accuracy and high efficiency,and has the clinical significance of auxiliary diagnosis.
机构地区 上海理工大学医疗器械与食品学院 上海交通大学医学院附属新华医院普外科 上海市胆道疾病重点实验室 上海交通大学医学院附属仁济医院胆胰外科 吉林大学第一医院肝胆胰外一科 吉林大学中日联谊医院普外科 国家癌基因重点实验室
出处 《中国实用外科杂志》 CAS CSCD 北大核心 2021年第3期310-314,共5页 Chinese Journal of Practical Surgery
基金 国家自然科学基金资助项目(No.81801797,No.31701108,No.81701749) 上海市青年科技英才扬帆计划(No.17YF1411700)。
分类号 R6 [医药卫生—外科学]
