Numerous medical models based on the personalized medicine approach have been investigated to provide more efficient treatments and improved health-care service. Metagenomic data - the genomic samples of microbial communities - appear to be one of the most valu- able sources to test the hypotheses for these models. However, interpreting this source is hard due to its very high dimension. As a result, some visualization methods have been proposed to deal with metagenomic data. These methods are not only for representing the numerical data but also for leveraging deep learning algorithms on the generated images to improve the diagnosis. In this study, we present an approach that uses Growing Self-Organizing Maps to transform features of three species metagenomic datasets into images. Then, generated images are fetched into a Convolutional Neural Network to do disease prediction tasks. The proposed method produces promising performance compared to other visualization approaches.