基于深度学习语义分割技术的气胸肺萎陷程度自动化计算  

Research on Calculation of Lung Compression Degree in Pneumothorax Using Semantic Segmentation Based on Deep Learning

作　　者：罗帅 刘安杰 张兴涛 占梦军 刘猛 范飞[1] 周宇驰 刘长远 邓振华[1] LUO Shuai;LIU Anjie;ZHANG Xingtao;ZHAN Mengjun;LIU Meng;FAN Fei;ZHOU Yuchi;LIU Changyuan;DENG Zhenhua(West China School of Basic Medical Sciences&Forensic Medicine,Sichuan University,Chengdu 610041,China;College of Computer Science,Sichuan University,Chengdu 610041,China;School of Computer Science and Engineering,University of Electronic Science and Technology of China,Chengdu 611731,China;Xinzheng Forensic Institute,Yibin 644022,Sichuan,China)

机构地区：[1]四川大学华西基础医学与法医学院,成都610041 [2]四川大学计算机学院,成都610041 [3]电子科技大学计算机科学与工程学院,成都611731 [4]宜宾鑫正司法鉴定所,四川宜宾644022

出　　处：《刑事技术》2024年第5期472-479,共8页Forensic Science and Technology

基　　金：四川省自然科学基金项目青年基金项目(24NSFSC6731);上海市现场物证重点实验室开放课题(2023XCWZK03)。

摘　　要：在法医临床学鉴定中,当需要确定肺萎陷程度时,通过Mimics软件计算被认为金标准。然而,由于Mimics软件的操作复杂且耗时较长,一些法医工作者仍然倾向于使用目测法、三线法等传统方法进行计算,这种做法可能导致鉴定意见出现一定程度的误差。本研究基于深度学习语义分割技术开发了肺萎陷程度自动化计算模型,并与Mimics软件计算肺萎陷程度的结果比对,以探究深度学习在肺萎陷程度测算中的可行性与可靠性。本研究收集包含气胸诊断的42例DICOM格式CT影像数据,每例图像约350张,层厚1 mm,从中随机选取32例数据用于模型训练,人工标注1943张图像中胸廓内含气区域,另外10例数据由Mimics软件测量肺萎陷程度,用于验证模型训练效果。同时,选取5例气胸相关鉴定案例作为外部测试集,通过模型和Mimics软件重建两种方法计算肺萎陷程度,分析两种方法结果的相关性及计算误差。在验证集中,模型计算结果与人工方法的平均误差为2.4%,平均计算时间为60.04 s;在测试集中平均误差为4.4%。本研究构建的模型在气胸所引起的肺萎陷程度自动化测算中表现出潜在的应用价值,为法医临床学中对气胸所致的肺萎陷程度准确定量提供了可靠的技术支撑。Calculation of the degree of lung compression by Mimics software remains the“gold standard”.In the forensic sphere,due to the complexity of the Mimics software,many people do not utilize this method in forensic practice.They may calculate degree of lung compression by visual observation,represent the result of degree of lung compression by some slicer of CT.These factors will lead to inaccuracies of calculated results.The aim of this study is to develop a model for automatic calculation of lung compression degree based on deep learning semantic segmentation technology,and explore the feasibility of deep learning for lung compression measurement by comparing the results of automatic calculations with Mimics software.In this study,42 cases of the computed tomography(CT)data including pneumothorax diagnosis in DICOM format were collected each cases has about 350 images with a thickness of 1 mm.Among them,32 cases used for training and 10 cases used for validation.The air-containing regions of 1943 images were manually annotated.An additional fi ve chest CT cases were selected for external testing.The degree of lung compression was calculated by both the deep learning model and Mimics software,and the correlation between the results of the two methods and the calculation errors were analyzed.In the validation set,the average error between the deep learning model calculation results and the manual method was 2.4%,and the model processed an average of 356 per case with an average time of 60.04 s,while the average error in the test set was 4.4%.The aforementioned results lead to the following conclusions:The deep learning model constructed in this study has the potential to be applied in the automated measurement of the lung compression degree due to pneumothorax,which can provide a reference for the calculation of the lung compression degree due to pneumothorax in forensic practice.

关 键 词：法医临床学 深度学习 MIMICS软件 肺萎陷 三维重建 

分 类 号：DF795.1[医药卫生—法医学]