During an epidemic crisis, medical image analysis namely microscopic analyses are made to confirm or not the existence of the epidemic pathogen in suspected cases. Pathogen are all infectious agents such as a virus, bacterium, protozoa, prion etc. However, there is often a lack of specialists in the handling of microscopes, hence allowing the need to make the microscopic analysis abroad. This results in a considerable loss of time and in the meantime, the epidemic continues to spread. To save time in the analysis of samples, we propose to make the future microscopes more intelligent so that they will be able to indicate by themselves the existence or not of the pathogen of an epidemic in a sample. To have a smart microscope, we propose a methodology based on efficient Convolution Neural Network (CNN) architecture in order to classify epidemic pathogen with five deep learning phases: (1) Training dataset of provided images (2) CNN Training (3) Testing data preparation (4) CNN generated model on testing data and finally (5) Evaluation of images classified. The resulted classification process can be integrated in a mobile computing solution on future microscopes. CNN can improve the accuracy in pathogens diagnosis that are focused on hand-tuned feature extraction implying some human mistakes. For our study, we consider cholera and malaria epidemics for microscopic images classification with a relevant CNN, respectively Vibrio cholerae images and Plasmodium falciparum images. Image classification is the task of taking an input image and outputting a class or a probability of classes that best describes the image. Interesting results have been obtained from the CNN model generated achieving the classification accuracy of 94%, with 200 Vibrio cholera images and 200 Plasmodium falciparum images for training dataset and 80 images for testing data. Although this document addresses the classification of epidemic pathogen images using a CNN model, the underlying principles apply to the other fields of science and technology, because of its performance and its capability to handle more layers than the previous traditional neural networks.