The multidimensional nature of metagenomic data with numerous features poses challenges for direct observation and analysis. Although T-distributed stochastic neighbor embedding (t-SNE) can be a potential solution for visualizing such complex data, it is necessary to examine sensitive factors affecting the quality of such visualization for classification tasks. This study evaluates a combination of fine-tuned t-SNE-based visualizations and simple Convolution Neural Networks (CNNs) to predict diseases on such visualizations. Firstly, the data binning technique is applied to discretize continuous values of metagenomic features, including bacterial abundance, to integer numbers that can be appropriated to transform into color points. Next, t-SNE is leveraged to represent such features on a two-dimensional image, followed by image classification tasks to determine whether the sample affected the disease. We have evaluated crucial hyperparameters of generating visualization with t-SNE, including the perplexity, the maximum number of iterations to optimize t-SNE, the transparency rates, and colormaps. Extensive experiments in seven metagenomic datasets related to six diseases have validated the effectiveness of the fine-tuned visualization with t-SNE compared to several previous studies.