Chen Ziyang, Wang Zhuowei, Zhao Meng, Zhao Qin, Liang Xuehu, Li Jiajian, Song Xiaoyu
School of Computers, Guangdong University of Technology, Guangzhou, China.
School of Computer Science and Engineering, Tianjin University of Technology, Tianjin, China.
Front Neurosci. 2022 Aug 25;16:807085. doi: 10.3389/fnins.2022.807085. eCollection 2022.
Automatic identification of Alzheimer's Disease (AD) through magnetic resonance imaging (MRI) data can effectively assist to doctors diagnose and treat Alzheimer's. Current methods improve the accuracy of AD recognition, but they are insufficient to address the challenge of small interclass and large intraclass differences. Some studies attempt to embed patch-level structure in neural networks which enhance pathologic details, but the enormous size and time complexity render these methods unfavorable. Furthermore, several self-attention mechanisms fail to provide contextual information to represent discriminative regions, which limits the performance of these classifiers. In addition, the current loss function is adversely affected by outliers of class imbalance and may fall into local optimal values. Therefore, we propose a 3D Residual RepVGG Attention network (ResRepANet) stacked with several lightweight blocks to identify the MRI of brain disease, which can also trade off accuracy and flexibility. Specifically, we propose a Non-local Context Spatial Attention block (NCSA) and embed it in our proposed ResRepANet, which aggregates global contextual information in spatial features to improve semantic relevance in discriminative regions. In addition, in order to reduce the influence of outliers, we propose a Gradient Density Multiple-weighting Mechanism (GDMM) to automatically adjust the weights of each MRI image a normalizing gradient norm. Experiments are conducted on datasets from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Australian Imaging, Biomarker and Lifestyle Flagship Study of Aging (AIBL). Experiments on both datasets show that the accuracy, sensitivity, specificity, and Area Under the Curve are consistently better than for state-of-the-art methods.
通过磁共振成像（MRI）数据自动识别阿尔茨海默病（AD）可以有效地辅助医生诊断和治疗阿尔茨海默病。当前的方法提高了AD识别的准确性，但它们不足以应对类间差异小和类内差异大的挑战。一些研究试图在神经网络中嵌入补丁级结构以增强病理细节，但巨大的规模和时间复杂性使这些方法并不理想。此外，几种自注意力机制未能提供上下文信息来表示判别区域，这限制了这些分类器的性能。此外，当前的损失函数受到类不平衡异常值的不利影响，可能会陷入局部最优值。因此，我们提出了一种堆叠有几个轻量级块的3D残差RepVGG注意力网络（ResRepANet）来识别脑部疾病的MRI，它还可以在准确性和灵活性之间进行权衡。具体来说，我们提出了一种非局部上下文空间注意力块（NCSA）并将其嵌入到我们提出的ResRepANet中，该块在空间特征中聚合全局上下文信息以提高判别区域中的语义相关性。此外，为了减少异常值的影响，我们提出了一种梯度密度多重加权机制（GDMM），通过归一化梯度范数自动调整每个MRI图像的权重。在来自阿尔茨海默病神经成像倡议（ADNI）和澳大利亚衰老成像、生物标志物和生活方式旗舰研究（AIBL）的数据集上进行了实验。两个数据集上的实验表明，准确率、灵敏度、特异性和曲线下面积始终优于现有方法。
