Code review, an inspection of code changes in order to identify and fix defects before integration, is essential in Software Quality Assurance (SQA). Code review is a time-consuming task since the reviewers need to understand, analysis and provide comments manually. To alleviate the burden of reviewers, automatic code review is needed. However, this task has not been well studied before. To bridge this research gap, in this paper, we formalize automatic code review as a multi-instance learning task that each change consisting of multiple hunks is regarded as a bag, and each hunk is described as an instance. We propose a novel deep learning model named DeepReview based on Convolutional Neural Network (CNN), which is an end-to-end model that learns feature representation to predict whether one change is approved or rejected. Experimental results on open source projects show that DeepReview is effective in automatic code review tasks. In terms of F1 score and AUC, DeepReview outperforms the performance of traditional single-instance based model TFIDF-SVM and the state-of-the-art deep feature based model Deeper.