Guided multi-modal image genetics data feature analysis method

Abstract

According to the guide type multi-modal image genetics data feature analysis method, multi-modal image data analysis and gene image canonical correlation analysis of samples are considered at the same time. Weights are divided into modal consistency weights and modal specificity weights by adopting a weight decomposition method, the modal consistency weights represent information shared among modals, and the modal specificity weights represent unique information of single modals. In addition, prior information, namely labels of samples, is adopted, feature learning of multi-modal image data is guided by utilizing regression analysis, and canonical correlation analysis of the multi-modal data and gene data is used as a plurality of learning tasks by utilizing a multi-task learning framework in machine learning; utilizing useful information contained in multi-task learning helps each task to obtain a more accurate learner and find differences and relationships among the tasks. According to the method disclosed by the invention, feature selection and image genetics data correlation analysis can be effectively carried out.

Description

technical field [0001] The invention discloses a guided multimodal image genetics data feature analysis method. Background technique [0002] Alzheimer's disease (Alzheimer's disease, AD), also known as senile dementia, is a primary degenerative brain disease, clinically manifested as memory and language impairment, as well as personality and behavior changes, and is one of the important diseases that endanger the health of the elderly. one. AD is generally divided into two stages, one is the pre-dementia stage, medically known as mild cognitive impairment (Mildcognitive impairment, MCI), MCI is between normal aging and AD, mainly manifested as memory loss, short-term memory it is good. One is the stage of dementia, which is divided into mild, moderate and severe. The memory impairment continues to aggravate and seriously affects life. AD develops slowly and is a persistent high-level neurological deficit. Neurochemically, AD often manifests as damage to multiple neurotr...

AI Technical Summary

Problems solved by technology

[0005] CN112348088A discloses a cross-modal data registration method based on mutual information canonical correlation analysis. The method maximizes the typical correlation degree of two modals by passing two modal data through a multimodal encoder respectively, although the modal Mode unique and public features, but only applicable to two modes, not very applicable to multiple modes

CN111340103A discloses a feature layer fusion method based on graph embedding canonical correlation analysis. The method uses L21 norm regularization to realize the selection method of independent complementary features of a single modal feature space, and constructs a data similarity matrix graph to describe single modal feature space. The proximity relationship of sample points does not take into account the correlation between modal features

[0006] In short, the existing multimodal fusion feature selection methods cannot better consider the relationship between modal features, which directly affects the correlation analysis between image data and genetic data, and the accuracy needs to be further improved.

Images