Multi instance learning via deep CNN for multi-class recognition of Alzheimer's disease
Multi instance learning via deep CNN for multi-class recognition of Alzheimer's disease
复制标题
DOI：
10.1109/iwcia47330.2019.8955006
复制
发表时间：
2019-11
期刊：
影响因子：
--
通讯作者：
M. Kavitha;N. Yudistira;Takio Kurita
中科院分区：
文献类型：
--
作者：
M. Kavitha;N. Yudistira;Takio Kurita
近年来，人们开发了多种阿尔茨海默病 (AD) 分类技术，这些技术基于手工制作的机器学习和晦涩的深度学习模型。本研究提出了一种基于类 Unet 2D 卷积神经网络（CNN）和多项逻辑回归分类器相结合的新分类框架，该框架在将 3D 正电子发射断层扫描（PET）图像选择为 2D 切片序列后学习切片内进行多类分类。 CNN 用于生成大脑的注意力特征，而逻辑回归则用于学习 AD 分类的各个类别的特定局部特征。在网络末端，我们在 softmax 之前使用平均池化层来解决四类分类问题。它可以有效地生成一类灵活的转换，并且可以通过反向传播进行端到端训练。结果表明，所提出的多实例学习（MIL）可以学习感兴趣区域（ROI）本身，因此可以帮助有效地识别 AD 的精确模式。所提出的将类 Unet CNN 与多项回归分类器相结合的方法在 AD 和 MCI 分类上分别达到了 97.9% 和 96.7% 的最高准确率。它远远高于文献中传统方法的性能。
In recent years, number of classification techniques for Alzheimer's disease (AD) have been developed that produced methods based on the use of hand-crafted machine learning and obscure deep learning models. This study proposed a new classification framework based on the combination of Unet-like 2D convolutional neural networks (CNN) and multinomial logistic regression classifier, which learns the intra-slice for multi-class classification after the selection of the 3D positron emission tomography (PET) image into a sequence of 2D slices. The CNNs are performed to generate the attention features of the brain while the logistic regression incorporated to learn those specifically localized features of various classes for AD classification. At the end of the network, we used a average pooling layer before the softmax for four-class classification problem. It can efficiently generate a flexible class of transformations and that can be trained end-to-end by back propagation. The results indicated that the proposed multi-instance learning (MIL) learns region of interest (ROI) itself and thus that could help to efficiently identify the precise patterns for AD. The proposed combined Unet-like CNN with multinomial regression classifier approach achieved highest accuracy of 97.9% and 96.7% on the classification of AD and MCI, respectively. It is much higher than the performances of the conventional methods in the literature.