Unsupervised deep frequency-channel attention factorization to non-linear feature extraction: A case study of identification and functional connectivity interpretation of Parkinson's disease

In the realm of contemporary medical research, the extraction of pivotal attributes associated with Parkinson's disease (PD) from intricate Magnetic Resonance Imaging (MRI) data characterized by multifaceted dimensions and non-linearity stands as a formidable challenge in the pursuit of realizing automated diagnostic support. To address this issue, an approach is introduced, denoted as the Deep Frequency-Channel Attention Factorization (Deep FCAF), designed to address the complexities inherent to the neurodegenerative ailment of Parkinson's. The Deep FCAF amalgamates three key elements: (1) the inherent capacity for non-linear processing exhibited by neural networks, (2) the inherent capability of the attention mechanism to capture global interdependencies, and (3) the underpinning principles of tensor decomposition within the order of multi-dimensional data structures. The experiments conducted on the identification of PD underscore the efficacy of Deep FCAF in assimilating structural and non-linear factor matrices, consequently leading to an improved classification performance. Notably, the elucidation of functional connectivity based on the acquired factor matrices align with previous scholarly investigations. The mathematical formulations, derived from the MRI data, facilitate the measurement of the dynamic mechanisms inherent in the MRI data.