BASIRA lab, Faculty of Computer and Informatics, Istanbul Technical University, Istanbul, Turkey; Université de Sousse, Ecole Nationale d'Ingénieurs de Sousse, LATIS- Laboratory of Advanced Technology and Intelligent Systems, 4023, Sousse, Tunisie.
Université de Sousse, Ecole Nationale d'Ingénieurs de Sousse, LATIS- Laboratory of Advanced Technology and Intelligent Systems, 4023, Sousse, Tunisie.
Med Image Anal. 2020 Oct;65:101768. doi: 10.1016/j.media.2020.101768. Epub 2020 Jun 27.
Existing graph analysis techniques generally focus on decreasing the dimensionality of graph data (i.e., removing nodes, edges, or both) in diverse predictive learning tasks in pattern recognition, computer vision, and medical data analysis such as dimensionality reduction, filtering and embedding techniques. However, graph super-resolution is strikingly lacking, i.e., the concept of super-resolving low-resolution (LR) graphs with n nodes into high-resolution graphs (HR) with [Formula: see text] nodes. Particularly, learning how to automatically generate HR brain connectomes, without resorting to the computationally expensive MRI processing steps such as image registration and parcellation, remains unexplored. To fill this gap, we propose the first technique to super-resolve undirected fully connected graphs with application to brain connectomes. First, we root our brain graph super-resolution (BGSR) framework in learning how to estimate a centered LR population-based brain graph representation, coined as connectional brain template (CBT), acting as a proxy in the target BGSR task. Specifically, we hypothesize that the estimation of a well-representative and centered CBT would help better capture the individuality of each LR brain graph via its residual distance from the population-based CBT. This will eventually allow an accurate identification of the most similar individual graphs to a new testing graph in the LR domain for the target prediction task. Second, we leverage the estimated LR CBT (i.e., population mean) to derive residual LR brain graphs, capturing the deviation of all subjects from the estimated CBT. Third, we learn multi-topology LR graph manifolds using different graph topological measurements (e.g., degree, closeness, betweenness) by estimating residual LR similarity matrices modeling the relationship between pairs of residual LR graphs. These are then fused so we can effectively identify for each testing LR subject its most K similar training LR graphs. Last, the missing testing HR graph is predicted by averaging the HR graphs of the K selected training subjects. Predicted HR from LR functional brain graphs boosted classification results for autistic subjects by 16.48% compared with LR functional graphs.
现有的图分析技术通常侧重于在模式识别、计算机视觉和医学数据分析等不同的预测学习任务中降低图数据的维数（即删除节点、边或两者），例如降维、过滤和嵌入技术。然而，图超分辨率技术却明显缺乏，即从概念上把具有 n 个节点的低分辨率（LR）图超分辨率到具有 [公式：见正文] 个节点的高分辨率（HR）图。特别是，如何学习自动生成 HR 脑连接组图，而不依赖于计算成本高昂的 MRI 处理步骤，如图像配准和分割，仍然是一个未被探索的问题。为了填补这一空白，我们提出了第一种用于无向全连通图的超分辨率技术，并将其应用于脑连接组图。首先，我们将脑图超分辨率（BGSR）框架建立在学习如何估计基于人群的 LR 脑图表示的基础上，这种表示被称为连接性脑模板（CBT），在目标 BGSR 任务中充当代理。具体来说，我们假设，估计一个具有代表性和中心化的 CBT 将有助于通过其与基于人群的 CBT 的残差距离更好地捕捉每个 LR 脑图的个体性。这最终将允许在 LR 域中准确识别与目标预测任务的新测试图最相似的个体图。其次，我们利用估计的 LR CBT（即群体均值）来获得残差 LR 脑图，捕捉所有受试者与估计 CBT 的偏差。第三，我们通过估计残差 LR 相似性矩阵来学习使用不同图拓扑度量（例如，度、接近度、介数）的多拓扑 LR 图流形，这些相似性矩阵用于建模残差 LR 图之间的关系。然后对这些流形进行融合，以便我们可以为每个测试 LR 受试者有效地识别其 K 个最相似的训练 LR 图。最后，通过对 K 个选定的训练受试者的 HR 图进行平均，来预测缺失的测试 HR 图。与 LR 功能脑图相比，LR 功能脑图预测自闭症患者的分类结果提高了 16.48%。
