Diabetes mellitus is seen as a disease which affects highly social, human and financial expenses for a nation. One of the most important reasons for creating machine learning systems is that in many areas experience is insufficient, and the codification of the knowledge that describes it is limited, fragmented and, therefore, incomplete. Artificial intelligence has made significant advances in fields such as education, agriculture and healthcare, where it can detect and treat diseases like cancer and diabetes long before traditional methods.The first phase of this study will present a comparative neural network analysis and will implement a neural network classifier optimized by Firefly algorithm to predict diabetes diagnosis based on factors mentioned in patients.The second phase of this work will develop a multi-scale convolutional neural network (MCNN) model for the early diagnosis of diabetes mellitus. The detection technique is based on a proposed tanning model by analyzing data from diabetic and non-diabetic patients from the PIMA Indian database.