基于深度学习的MRI图像自动分割技术在胎盘植入性疾病预测中的应用  

Application of automatic segmentation technology of MRI images based on deep learning inthe prediction of placenta accreta disease

作　　者：李倩 廖蔚 戢兰蝶 叶莉丽 梅劼 LI Qian;LIAO Wei;JI Lan-die;YE Li-li;MEI Jie(Clinical Medical School,Southwest Medical University,Luzhou 646000,China;Department of Obstetrics,Sichuan Academy of Medical Sciences·Sichuan Provincial People's Hospital(Affiliated Hospital of University of Electronic Science and Technology of China),Chengdu 610072,China;Clinical Medical School of University of Electronic Science and Technology,Chengdu 610072,China)

机构地区：[1]西南医科大学临床医学院,四川泸州646000 [2]四川省医学科学院·四川省人民医院(电子科技大学附属医院)产科,四川成都610072 [3]电子科技大学临床医学院,四川成都610072

出　　处：《实用医院临床杂志》2024年第6期120-123,共4页Practical Journal of Clinical Medicine

基　　金：四川省科技厅重点研发项目(编号:2023YFS0039)。

摘　　要：目的探讨基于深度学习的MRI图像自动分割技术在胎盘植入性疾病(PAS)预测中的作用。方法收集2016年1月至2022年5月四川省人民医院孕晚期胎盘MRI图像。训练集:合并PAS的患者58例760张图像;未合并PAS的患者41例596张图像。测试集:合并PAS的患者18例232张图像;未合并PAS的患者11例161张图像。训练完成后分别与两名影像科主治医师及住院医师诊断结果进行对比。结果在PAS的二分类预测中:Nasnet神经网络分类模型预测PAS发生的灵敏度为100%,特异度为90.9%,准确性为96.5%,AUC为0.985。住院医师组灵敏度72%、特异度63.6%,准确率68.9%;主治医师组灵敏度88.8%,特异度81.8%,准确性86.2%。住院医师组预测准确率与分类模型对比,差异有统计学意义(P<0.05),二分类模型预测PAS的准确率远胜于住院医师组,二者灵敏度、特异度比较,差异无统计学意义(P>0.05)。主治医师组预测效果与二分类模型预测效果对比,差异无统计学意义(P>0.05),但Kappa值为0.776,两者一致性较好。结论基于深度学习技术的MRI图像自动分割在预测PAS中是可行的。Objective To explore the role of automatic MRI image segmentation technology based on deep learning in the prediction of placenta accreta spectrum(PAS)disorders.Methods Late pregnancy placental MRI images from January 2016 to May 2022 in Sichuan Provincial People's Hospital were collected.The training set included 760 images of 58 patients with PAS and 596 images of 41 patients without PAS.The test set included 232 images of 18 patients with PAS and 161 images of 11 patients without PAS.After the training,the diagnostic results were compared with those of attending physicians and residents at the department of imaging.Results The sensitivity of the Nasnet neural network in the binary classification model was 100%for predicting the occurrence of PAS.The specificity of the model was 90.9%.The accuracy was 96.5%and AUC was 0.986.In predicting the occurrence of PAS,the sensitivity of the residents group was 72.0%,the specificity was 63.6%and the accuracy was 68.9%.The sensitivity of the attending physicians group was 88.8%,the specificity was 81.8%and the accuracy was 86.2%.There was significant difference when comparing the accuracy between the resident group and the classification model(P<0.05).Thus,the accuracy of the binary classification model to predict PAS was much better than that of the resident group.However,there was no significant difference in sensitivity and specificity between the resident group and the binary classification model(P>0.05).There was no significant difference in the prediction effect between the attending physician group and the binary classification model(P>0.05).Moreover,Kappa value was 0.776,suggesting a great agreement between the two methods.Conclusions Conclusion Automatic MRI image segmentation based on deep learning technology is feasible in predicting PAS.

关 键 词：胎盘植入性疾病 深度学习 核磁共振 人工智能 风险预测 

分 类 号：R71[医药卫生—妇产科学]