Accurate localization of all root canal orifices is critical for the success of root canal treatment. However, concealed orifices, particularly the second mesiobuccal canal (MB2), are frequently missed due to ambiguous visual cues, leading to high rates of treatment failure. To address this challenge, we propose a novel computer-aided detection framework, OrificeNet, the first to perform orifice detection directly from intraoperative microscope-view RGB images. Our framework formulates this as a segmentation task, employing an encoder-decoder network with a multi-scale strategy and a hierarchical cascaded decoder to effectively identify orifices. Furthermore, to simulate the real clinical workflow, we introduce a CBCT-guided post-processing step that leverages pre-operative 3D data to refine the 2D prediction via an affine transformation, accurately locating and completing concealed orifices. Extensive experiments on a clinically collected dataset demonstrate that our proposed method significantly achieve better performance. Our work presents an effective and clinically-translatable solution to reduce the risk of missed canals, enhancing the success rate of root canal treatments.