摘要
目的:探讨在机器深度学习中迁移学习在图像软件对实验动物解剖结构的识别、提取和自动分割中的作用,以及数据增强算法对迁移学习能力的影响。方法:在HyVision Ablation Planning V1.0图像软件的平台上,以Efficient Net b1神经网络作为深度学习的骨干网络。利用51套VX2兔肝癌模型的计算机断层扫描(computed tomography,CT)图像,以数据增强的方式进行迁移学习训练。将图像软件已经具备的人体腹部CT图像上器官的识别、提取与自动分割功能在动物模型上进行重现。比较不同的学习模型和算法模型的Dice系数、归一化表面Dice(normalized surface Dice,NSD)、三维重建的图像质量及与医师标注的动物模型训练集的差异。结果:从有数据增强无迁移学习型的模型到有数据增强有迁移学习型的模型,VX2兔CT图像的器官自动分割Dice系数从0.525提升到0.676,提高了28.76%,NSD从0.448提升到0.616,提高了37.50%。从无数据增强有迁移学习型的模型到有数据增强有迁移学习型的模型,VX2兔CT图像的器官自动分割Dice系数从0.502提升到0.676,提高了34.66%,NSD从0.459提升到0.616,提高了34.20%。表明在机器深度学习过程中迁移学习与数据增强对于研究新的解剖对象同等重要。结论:在机器深度学习过程中,迁移学习的功能可以借助数据增强算法,获得更好的图像识别、提取与自动分割的结果。
Objective:To explore the role of transfer learning in the recognition,extraction,and automatic segmentation of anatomical structures in graphic software during machine deep learning,and the impact of data augmentation algorithms on the ability of transfer learning.Method:The Efficient Net b1 neural network is used as the backbone network for deep learning on the platform of HyVision Ablation Planning V1.0 graphic software.51 sets of VX2 rabbit liver cancer model computed tomography(CT)scan images are used for transfer learning training through data augmentation.The graphic software's existing functions of identifying,extracting,and automatically segmenting organs on human abdominal CT images are reproduced on animal models.The differences in Dice and normalized surface Dice(NSD)values under different learning models,as well as the differences between the image quality of 3D reconstruction and the doctor-marked animal model training set,are compared.Result:From the model with"data augmentation without transfer learning"to the model with"data augmentation with transfer learning",the Dice index for automatic organ segmentation of VX2 rabbit CT scan images increased from 0.525 to 0.676,an increase of 28.76%,and the NSD index increased from 0.448 to 0.616,an increase of 37.50%.From the model with"no data augmentation but with transfer learning"to the model with"data augmentation with transfer learning",the Dice index for automatic organ segmentation of VX2 rabbit CT scan images increased from 0.502 to 0.676,an increase of 34.66%,and the NSD index increased from 0.459 to 0.616,an increase of 34.20%.This indicates that in the process of machine deep learning if studying new anatomical objects,transfer learning and data augmentation have equally important roles.Conclusion:In the process of machine deep learning,the function of transfer learning can obtain better graphic recognition,extraction,and automatic segmentation results with the help of"data augmentation"algorithm.
作者
张雨萌
康文迪
席俊青
池琛
杨正强
Zhang Yumeng;Kang Wendi;Xi Junqing;Chi Chen;Yang Zhengqiang(Department of Interventional Therapy,National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital,Chinese Academy of Medical Sciences&Peking Union Medical College,Beijing 100021,China)
出处
《肝癌电子杂志》
2023年第3期52-57,共6页
Electronic Journal of Liver Tumor
基金
国家重点研发计划(2020YFC0122300)。
关键词
迁移学习
数据增强
动物CT图像分割
神经网络
Transfer learning
Data augmentation
Animal CT image segmentation
Neural network
