It is important to segment brain tumors to diagnose, treat, and monitor them using CT images. The traditional U-Nets are not expected to generate complicated anatomy and tight tumor boundaries. In this paper we offer an improved Attention U-Net model that can better segment brain tumors. Attention also assists the network in targeting the tumor areas as there is fewer background noise. After being trained and validated on a set of CT scans of brain tumors our model outperformed the baseline U-Net regarding Dice and IoU scores. The issue of attention enhances definition of boundaries as well as representation of features. This work is a promising development towards computer aided diagnosis in clinics.