基于腹部平扫CT深度学习对脾脏损伤区域分割及严重程度预测  

Splenic injury region segmentation and severity prediction based on deep learning of abdominal plain CT

作　　者：周里平 罗德 李波 程智涛 冷妍霖 唐泳 刘燚 白昌松 古广强 苏松 Zhou Liping;Luo De;Li Bo;Cheng Zhitao;Leng Yanlin;Tang Yong;Liu Yi;Bai Changsong;Gu Guangqiang;Su Song(Department of General Surgery(Hepatobiliary Surgery),the Affiliated Hospital of Southwest Medical University,Academician(Expert)Workstation of Sichuan Province,Metabolic Hepatobiliary and Pancreatic Diseases Key Laboratory of Luzhou City,Luzhou,Sichuan Province 646000,China;School of Computer Science and Engineering,University of Electronic Science and Technology of China,Chengdu 611731,China;School of Information and Software Engineering,University of Electronic Science and Technology of China,Chengdu 610054,China;Institute of Intelligent Chinese Medicine,Chongqing College of Traditional Chinese Medicine,Chongqing 402760,China;Southwest Medical University,School of Basic Medicine(Clinical Medicine,Class of 2022),Luzhou,Sichuan Province 646000,China;General Surgery Department of Xuyong County People’s Hospital,Xuyong,Sichuan Province 646400,China;Department of Hepatobiliary and Pancreatic Surgery,The Third People’s Hospital of Mianyang,Sichuan Mental Health Center,Mianyang,Sichuan Province 621000,China)

机构地区：[1]西南医科大学附属医院普通外科(肝胆外科),西南医科大学附属医院四川省院士(专家)工作站,代谢性肝胆胰疾病泸州重点实验室,四川泸州646000 [2]电子科技大学,计算机科学与工程学院,成都611731 [3]电子科技大学,信息软件工程学院,成都610054 [4]重庆中医药学院,数智中医药创新研究院,重庆402760 [5]西南医科大学,基础医学院(2022级临床医学),四川泸州646000 [6]叙永县人民医院普外科,四川叙永646400 [7]四川省精神卫生中心,绵阳市第三人民医院肝胆胰外科,四川绵阳621000

出　　处：《创伤外科杂志》2025年第3期177-185,共9页Journal of Traumatic Surgery

基　　金：四川省科技厅中央引导地方科技发展专项项目(2024ZYD0270);叙永县人民医院-西南医科大学科技战略合作项目(2023XYXNYD07)。

摘　　要：目的构建基于腹部平扫CT的深度学习模型,探索其自动分割腹部闭合性损伤中脾脏损伤区域及初步评估脾脏损伤严重程度的价值。方法回顾性收集腹部平扫CT诊断为腹部闭合性损伤中的脾脏损伤患者122例,以及正常完整脾脏人员240例,共4290张腹部平扫CT影像。其中男性192例,女性170例;年龄5~90岁,平均53.2岁。在自动分割脾脏损伤区域的研究中,使用UNet、FCN-ResNet50、FCN-ResNet101三种深度学习模型。分别采用两步法(先分割脾脏轮廓再识别损伤区域)和一步法(直接识别损伤区域)进行深度学习模型的构建,并使用准确率(ACC)、交并比(IoU)以及Dice系数作为评价指标进行比较。在评估脾脏损伤严重程度的研究中,基于美国创伤协会(AAST)器官损伤量表(OIS),使用DenseNet、CNN、ResNet101、ResNet18四种深度学习模型。分别在完整的CT横断面图像和自动分割后的脾脏损伤区域上,对脾脏损伤进行轻症组(Ⅰ、Ⅱ级)和中重症组(Ⅲ~Ⅴ级)的二分类,并以受试者操作特征(ROC)曲线的曲线下面积(AUC)及其95%CI、ACC、灵敏度以及特异度来评估模型的分类性能。结果采用一步法完成脾脏损伤区域分割的模型FCN-ResNet101对脾脏损伤区域分割准确率最高(IoU:0.713,ACC:0.887,Dice系数:0.801)。采用直接输入完整图像识别的模型ResNet18对脾脏损伤严重程度预测的综合表现最佳(AUC:0.834,95%CI:0.607~1.000,ACC:0.770,灵敏度:0.885,特异度:0.561)。结论基于腹部平扫CT的深度学习模型FCN-ResNet101,对脾脏损伤具有较好的自动分割性能。在评估脾脏损伤严重程度时,模型ResNet18具有较好的分类性能。两模型均具备较好的临床应用前景。Objective To construct a deep learning model based on abdominal CT plain scans and to explore its value in automatically segmenting the region of splenic injury in closed abdominal injuries and initially assessing the severity of splenic injury.Methods A total of 4,290 abdominal plain CT images were retrospectively collected from 122 patients with splenic injuries diagnosed as closed abdominal injuries and 240 patients with normal intact spleens,of whom 192 were male patients and 170 were female,with a mean age of 53.2(5-90)years.Three deep learning models,UNet,FCN-ResNet50,and FCN-ResNet101,were used in the study of automatic segmentation of splenic injury regions.The two-step method(segmenting the spleen contour before recognizing the damaged region)and the one-step method(directly recognizing the damaged region)were used for the construction of the deep learning model,respectively,and the accuracy(ACC),intersection over union(IoU),and the DICE coefficient(Dice)were used as evaluation metrics for comparison.In the study to assess the severity of spleen injury,four deep learning models,DenseNet,CNN,ResNet101,and ResNet18,were used.Splenic injuries were dichotomized into mild(gradesⅠandⅡ)and moderately severe(gradesⅢ-Ⅴ)groups on complete CT cross-sectional images and automatically segmented splenic injury regions,respectively,and the classification performance of the models was evaluated by the area under the curve(AUC)of the receiver operating characteristic curves(ROC curves)and their 95%CI,ACC,sensitivities,and specificities.Results The model FCN-ResNet101,which uses the one-step method approach to complete splenic injury region segmentation,had the highest ACC for splenic injury region segmentation(IoU:0.713,ACC:0.887,Dice:0.801).The model ResNet18,which used direct input for complete image recognition,had the best overall performance for spleen injury severity prediction(AUC:0.834,95%CI:0.607-1.000,ACC:0.770,Sensitivity:0.885,Specificity:0.561).Conclusion FCN-ResNet101,a deep learning model based on

关 键 词：脾脏损伤 腹部闭合性损伤 平扫CT 

分 类 号：R657.6[医药卫生—外科学] R445[医药卫生—临床医学]