Abstract: Fundus image microaneurysm is the initial symptom of diabetic retinopathy. Automatic detection of microaneurysm based on color fundus image is helpful to assist doctors in determining whether the patient's retina is normal, and it is also the most important pre-treatment method in the grading evaluation of diabetic retinopathy. Because of the complex structure of the retina and the different brightness and contrast existed in the imaging of fundus images due to different factors such as patients, environment and collection devices, the existing microaneurysm detection algorithm is difficult to realize the accurate detection of microaneurysm, and test results include a large amounts of the microaneurysm candidate section, such as blood vessels, background noise. Convolutional neural network has very strong expression ability, it can automatically learn the characteristics of the target by training model, and thus this paper puts forward the microaneurysm detection based on convolutional neural network method. The simulation results show that the proposed method is superior to the traditional test method of microaneurysm, the accurate extraction of microaneurysm can be realized under the complex diabetic retinopathy, but some blood vessels and background noise were excluded. The number of candidate regions extracted based on convolutional neural network is less and the form is regular, which is conducive to subsequent feature extraction and classification.