多尺度互学习的三元组图卷积神经网络模拟的数据集

利用尺度不同的多个模板将大脑进行粗细不同的区域划分，然后计算相关的结构或功能连接并利用K近邻构造样本对应的图。在同一尺度下，我们利用三个样本图卷积进行有效距离的度量，再通过加权的方式融合不同尺度下的相似度度量结果。MTGCN算法不仅学习了不同尺度的脑连接网络的图表示，而且利用triplet损失函数挖掘更复杂的多样本关系，整合从粗到细的空间信息, 然后利用三元组图卷积网络为每个尺度下的连接网络进行被试的图表示学习。在注意缺陷多动障碍ADHD, ADNI, 白质脱髓鞘症等多疾病分类中提高3-8%的精度，验证了基于多尺度互学习的三元组图卷积神经网络的有效性。

Multiple templates with distinct scales are utilized to parcellate the brain into regions of varying granularity. Next, associated structural or functional connections are computed, and sample-specific graphs are constructed via k-nearest neighbors (KNN). At each individual scale, we adopt three sample-based graph convolutions to conduct effective distance measurement, then fuse the similarity measurement outputs across different scales using a weighted fusion scheme. The MTGCN algorithm not only learns graph representations of brain connectivity networks across different scales, but also leverages the triplet loss function to explore more complex multi-sample relationships and integrate coarse-to-fine spatial information. Subsequently, it employs the triplet graph convolutional network to perform graph representation learning for study subjects based on the connectivity networks at each scale. When applied to multi-disease classification tasks including Attention Deficit Hyperactivity Disorder (ADHD), ADNI, and white matter demyelination, the proposed method achieved a 3-8% improvement in classification accuracy, which validates the effectiveness of the triplet graph convolutional neural network based on multi-scale mutual learning.

搜集汇总

背景与挑战

背景概述

该数据集用于支持多尺度互学习的三元组图卷积神经网络（MTGCN）的模拟，通过多尺度模板划分大脑区域并构建图结构，以学习脑连接网络的图表示。它旨在提高注意力缺陷多动障碍（ADHD）、ADNI和白质脱髓鞘疾病等多疾病分类的准确性，应用领域涉及脑成像和机器学习。