Thyroid classification via new multi-channel feature association and learning from multi-modality MRI images
Thyroid classification via new multi-channel feature association and learning from multi-modality MRI images
复制标题
通过新的多通道特征关联和多模态 MRI 图像学习进行甲状腺分类
DOI：
10.1109/isbi.2018.8363573
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
D. Feng
中科院分区：
文献类型：
--
作者：
Rong Zhang;Qiufang Liu;Hui Cui;Xiuying Wang;Shaoli Song;Gang Huang;D. Feng
甲状腺组织的计算机分类对于甲状腺患者的精确诊断和治疗计划至关重要。在这项工作中，我们提出了一种新的多通道特征关联和融合学习（FAFL）模型，用于多模态MRI图像中的甲状腺组织分类。我们的模型有三层，首先是一个两层卷积神经网络（CNN），用于产生三个多通道张量;其次是一个多特征关联层，用于融合来自三个CNN张量的相应多模态特征，以生成一个新的特征关联张量;第三，一个级联层，用于完全连接多通道CNN张量和新的特征关联张量进行分类。我们的模型已经在45项患者研究中进行了训练和验证。实验结果表明，我们的FAFL模型优于传统的CNN与单模态图像作为输入，以及两个模型，考虑三种模态，但在分类精度，特异性和灵敏度方面的不同的关联方式。
Computerized classification of thyroid tissues is essential towards precision diagnosis and treatment planning for thyroid patients. In this work, we propose a novel multichannel feature association and fusion learning (FAFL) model for thyroid tissue classification from multi-modality MRI images. Our model has three layers including firstly a two-layer convolutional neural network (CNN) for producing three multi-channel tensors; secondly, a multifeature association layer to fuse corresponding multi-modality features from the three CNN tensors to generate a new feature association tensor; and thirdly, a concatenation layer to fully connect the multi-channel CNN tensors and the new feature association tensor for classification. Our model has been trained and validated on 45 patient studies. The experimental results demonstrated that our FAFL model outperformed the conventional CNN with single modality images as inputs as well as two models which consider three modalities but with different association ways in terms of classification accuracy, specificity and sensitivity.