Early detection and intervention has the potential for slowing disease progression of dementia. Brain segmentation of T1W structural MRI is an effective biomarker for assessing cognitive decline. However, automatic brain segmentation has required lengthy processing time and utilizes single structure for risk stratification. In this study, we developed a 3D fully convolutional neural network for ultrafast brain segmentation. Both qualitative and quantitative analysis demonstrated that our segmentation method has a strong generalization capability achieving promising experimental results. Furthermore, we utilized multiple regions in combination and defined a new biomarker to better differentiate early disease progression of normal versus MCI in addition to MCI versus Dementia. Our multi-region approach outperforms conventional single biomarkers.