| 李杰文,叶俊勇,徐胜生,伍文丽.特征迁移结合主动标注的颅内病变鉴别方法[J].电子测量与仪器学报,2021,35(9):186-194 |
| 特征迁移结合主动标注的颅内病变鉴别方法 |
| Identification of intracranial lesions based onfeature transfer and active labeling |
| |
| DOI： |
| 中文关键词: 脑炎 胶质瘤 特征迁移 主动学习 熵不确定性 |
| 英文关键词:encephalitis glioma feature transfer active learning entropy uncertainty |
| 基金项目:中央高校基本科研业务费(2018CDXYGD0017)、重庆市基础研究与前沿探索专项(cstc2018jcyjAX0633)项目资助 |
|
|
|
|
| 摘要点击次数: 202 |
| 全文下载次数: 678 |
| 中文摘要: |
| 为了解决使用磁共振图像进行脑炎和胶质瘤的临床诊断时会出现误诊的问题,提出了一种卷积层特征迁移结合主动样
本标注的分类方法。 该方法首先采用卷积层特征参数迁移,利用磁共振成像(MRI)数据进行模型的微调,并验证了不同 MRI
模态特征的区分能力;其次针对样本标注困难问题,设计基于熵不确定性的样本主动标注算法,提取样本的不确定性信息,进一
步提高模型的收敛速度以及泛化能力。 在由重庆医科大学附属第一医院放射科纳入的 175 个病例(脑炎 118 例,胶质瘤 57 例)
上进行实验,在交叉验证下分类准确率达到 95. 08%, 曲线下面积达到 0. 98,模型的分类精度显著优于现阶段主要依靠医生经
验的方法,准确率和曲线下面积分别提高 17. 51%和 0. 15;同时仅需要标注 30%的数据样本,模型便能达到最优性能,减少大量
数据标注工作,能够为初期诊断提供有意义的指导。 |
| 英文摘要: |
| In order to solve the misdiagnosis of encephalitis and glioma in clinical diagnosis while using MRI images, we proposed a
classification method of convolutional layer feature transfer combined with active sample labeling. The method firstly adopts the
convolutional layer features parameter transfer and uses the multi-modal MRI image data for the fine-tuning of models to verify the
distinguishing ability of different MRI modal features. Secondly, in view of the difficulty of sample labeling, an entropy uncertainty based
active labeling algorithm is designed to extract the uncertainty information of samples to further improve the convergence speed and
generalization ability of the model. Experiments were carried out on a dataset of 175 cases (118 cases of encephalitis and 57 cases of
glioma) collected by the radiology department of the First Affiliated Hospital of Chongqing Medical University. The results show that the
classification accuracy under cross-validation reached 95. 08% and area under the curve reached 0. 98. The accuracy of the model was
superior to the method mainly relying on the experience of doctors at present; and the accuracy and area under the curve was 17. 51%
and 0. 15 higher than that of doctors, respectively. At the same time, only 30% of the data samples need to be annotated, so the model
can achieve optimal performance, reduce a lot of data annotation work, and provide meaningful guidance for the initial diagnosis. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
