TDCVT: A Tucker Decomposition Based Cubic Vision Transformer Framework for Early Cognitive Impairment Detection

Alzheimer's disease (AD) is a widespread degenerative neurological disorder. Timely diagnosis of early mild cognitive impairment (EMCI) is considered vital for controlling the development of AD. Diffusion tensor imaging (DTI), as the only technique capable of characterizing changes in brain microstructure in vivo, generates fractional anisotropy (FA) maps with three-dimensional spatial details. However, existing methods face issues such as feature redundancy, high computational complexity, and insufficient interpretability when processing high-dimensional FA data. To this end, we proposes a novel classification framework called TDCVT, which combines orthogonal constrained tensor Tucker Decomposition with Cubic Visual Transformer (3D-ViT) to achieve efficient classification of EMCI. Orthogonal constraint Tucker decomposition suppresses redundant features through dynamic rank selection, while 3DViT captures global dependencies through 3D patch partitioning and multi-head self-attention mechanism. The experiment was based on a self-built dataset (EMCI/CN=210/240) from ADNI database, and the classification results of TDCVT demonstrated an accuracy of 84.4 %, which was 8.8 % higher than other combinational methods. This study sheds light on the process of high-dimension tensor and provides a high-precision, computational efficient and interpretable solution for the diagnosis of EMCI.