Data visualization is still a challenge for numerous fields. For metagenomic data, datasets are usually characterized by very high-dimensional data which are hard to interpret to humans. Among diseases using metagenomic data for prediction, deep learning usually yields a lower performance comparing to classical machine learning for colorectal cancer prediction. In this paper, we present an approach using manifold learning with t-distributed stochastic neighbor embedding (t-SNE) and spectral embedding to visualize numerical data into images and leverage deep learning algorithms to improve the performance in colorectal cancer diseases prediction. The work also provides promising potentials to improve the visualization quality and performance in prediction tasks on dense data. The analytical results of samples coming from five various regions including America, China, Austria, Germany, and France show promising in use of combination between these visualization approaches and deep learning to enhance the performance in colorectal cancer disease diagnosis.