摘要
目的 评价不同模型及不同学习方法的人工智能算法在磁共振T1增强成像中对肛瘘内口诊断的准确性。方法 回顾性分析上海中医药大学附属岳阳中西医结合医院2019年5月至2021年5月58例肛瘘病人及45例正常病例的磁共振T1增强序列图像,通过数据增强的方法将图像扩增至3 400幅,根据是否患病进行分层,采用分层随机抽样的方法将数据分为训练组(n=2 720)和验证组(n=680)。采用迁移学习和端到端学习两种方式对两组病例进行学习及测试,对比分析ResNet-18、ResNet-34和DenseNet-121的网络模型性能,根据结果对网络模型进行评估,分析不同方法及模型诊断的灵敏度及特异度。结果 采用ResNet-34模型迁移学习方式对肛瘘内口诊断的灵敏度及特异度为分别为96.97%和94.94%,评价效果最优。结论 基于磁共振T1增强序列的ResNet-34模型迁移学习算法可对肛瘘内口进行有效诊断,有助于提高诊断效能。
Objective To evaluate the accuracy of different models and different learning methods of artificial intelligence algorithms in the diagnosis of the internal orifice of anal fistula based on magnetic resonance(MR) T1 enhanced imaging.Methods MR T1 enhanced sequence images of 58 anal fistula patients and 45 normal individuals in Yueyang Hospital of Integrated Traditional Chinese and Western Medicine, Shanghai University of Traditional Chinese Medicine from May 2019 to May 2021 were retrospectively analyzed. The number of images was enlarged to 3 400 by data augmentation method, stratified according to whether the disease was present, and the data were assigned into training group(n=2 720) and verification group(n=680) by stratified random sampling method. The ResNet-18, ResNet-34 and DenseNet-121 network models were both constructed for anal fistula. Transfer learning and end-to-end learning were utilized to train the above two models on the training set images. Then the testing set images were assigned to validate the models, comparing the performance of two kinds of models and analyzing the differences between two learning methods.Results The sensitivity and specificity of the Resnet-34 model via transfer learning method for the diagnosis of the internal orifice of anal fistula were 96.97% and 94.94%, respectively, which the evaluation effect was the best.Conclusion The ResNet-34 model transfer learning algorithm based on the MR T1 enhanced sequence can effectively diagnose the internal orifice of anal fistula and improve the diagnostic efficiency.
作者
袁军
陈欣悦
常时新
王乐
周自明
YUAN Jun;CHEN Xinyue;CHANG Shixin;WANG Le;ZHOU Ziming(Yueyang Hospital of Integrated Traditional Chinese and Western Medicine,Shanghai University of Traditional Chinese Medicine,Shanghai 200437,China;Xuhai College,China University of Mining and Technology,Xuzhou,Jiangsu 221008,China;Academy for Engineering and Technology,Fudan University,Shanghai 200433,China)
出处
《安徽医药》
CAS
2023年第3期447-452,共6页
Anhui Medical and Pharmaceutical Journal
基金
国家自然科学基金项目(62076070)。
关键词
直肠瘘
磁共振成像
神经网络
图像分类
图像处理
计算机辅助
Rectal fistula
Magnetic resonance imaging
Neural network
Image classification
Image processing,computer-assisted
