Accurate segmentation of mandibular canals in lower jaws is important in dental implantology, in which the implant position and dimensions are currently determined manually from 3D CT images by medical experts to avoid damaging the mandibular nerve inside the canal. In this paper, we propose a novel dual-stage deep learning based scheme for automatic detection of mandibular canal. Particularly, we first we enhance the CBCT scans by employing the novel histogram-based dynamic windowing scheme which improves the visibility of mandibular canals. After enhancement, we design 3D deeply supervised attention U-Net architecture for localize the volume of interest (VOI) which contains the mandibular canals (i.e., left and right canals). Finally, we employed the multi-scale input residual U-Net architecture (MS-R-UNet) to accurately segment the mandibular canals. The proposed method has been rigorously evaluated on 500 scans and results demonstrate that our technique out performs the existing state-of-the-art methods in term of segmentation performance as well as robustness.