Publication

Abstract : The growing multimodality of biomedical data is uncovering the intricate relationships between biological processes that are hidden from view. Most of the time, these non-linear collaborations are illustrated using deep learning (DL)-based information fusion algorithms. Thus, we evaluate the present status of the specialty of such procedures and propose a point-by-point scientific categorization that works with the improvement of new methods as well as a wiser decision of combination methodologies for organic applications. Thus, we see that deep fusion techniques generally beat shallow and unimodal ones. The suggested subcategories of fusion techniques also exhibit various benefits and disadvantages. Joint representation learning is the optimum method, especially for intermediate fusion strategies, according to a review of current methodologies. This is because it successfully captures the complex relationships between many levels of biological organization. We conclude by highlighting progressive fusion as a possible area for future research, depending on pre-existing biological knowledge or search strategies. Similarly, applying transfer learning could help multimodal datasets overcome sample size restrictions. At the point when these datasets become all the more openly accessible, multimodal DL techniques provide the opportunity to create comprehensive models that can identify the intricate administrative cycles underlying health and illness.