Minimal invasive mitral valve repair (MIVMR), a recently developed method guided by video endoscopy, still remains complex and technically challenging. The clinical outcomes are strongly dependent on the appropriate choice of the thoracic incision, since it enables, or not, a good visualization and manipulation of the valve. This makes the accurate localization of mitral valve very important. In this paper, we propose a mitral valve deep learning localization method employing an edgepoint-based approach instead of traditional segmentation. Leveraging existing annotations of heart chambers, our method extracts relevant 2D slices surrounding the mitral valve annulus from CT scans, capturing both global structural and local texture features. These fused features are processed through a deep learning-based 2D landmark localization network to automatically detect the leaflet edgepoints. Subsequently, mitral valve annulus can be generated by the interpolation of these edgepoints. Our method achieves the best result of 8.76° and 2.83 mm in terms of orientational and positional error on local cardiac CT dataset. Our approach automatically delineates the mitral valve's unique saddle-shaped contour and provides precise positional information, facilitating improved surgical planning.