Accurate segmentation of a set of regions of interest (ROIs) in the brain images is a key step in many neuroscience studies. Due to the complexity of image patterns, many learning-based segmentation methods have been proposed, including auto context model (ACM) that can capture highlevel contextual information for guiding segmentation. However, since current ACM can only handle one ROI at a time, neighboring ROIs have to be labeled separately with different ACMs that are trained independently without communicating each other. To address this, we enhance the current single-ROI learning ACM to multi-ROI learning ACM for joint labeling of multiple neighboring ROIs (called eACM). First, we extend current independently-trained single-ROI ACMs to a set of jointly-trained cross-ROI ACMs, by simultaneous training of ACMs for all spatially-connected ROIs to let them to share their respective intermediate outputs for coordinated labeling of each image point. Then, the context features in each ACM can capture the cross-ROI dependence information from the outputs of other ACMs that are designed for neighboring ROIs. Second, we upgrade the output labeling map of each ACM with the multi-scale representation, thus both local and global context information can be effectively used to increase the robustness in characterizing geometric relationship among neighboring ROIs. Third, we integrate ACM into a multi-atlases segmentation paradigm, for encompassing high variations among subjects. Experiments on Loni LPBA40 dataset show much better performance by our eACM, compared to the conventional ACM.