The precise detection of cephalometric landmarks on two-dimensional (2D) radiographs or three-dimensional (3D) computed tomography (CT) images is a fundamental step in various medical fields, especially in research on orthodontics and maxillofacial surgery. Deep learning-based detectors have demonstrated remarkable accuracy in 2D cephalometric analysis, whereas conventional single-view approaches are limited by their reliance on information from a single perspective. This study proposes GeoRefineNet, a novel multistage framework that leverages information from multiple CT scans acquired at various angles. By incorporating geometric knowledge through a 3D heatmap reconstruction process, GeoRefineNet improves robustness, accuracy, and adaptability to various cephalometric configurations. The proposed framework predicts 3D landmark positions on CT images, effectively addressing challenges associated with high-dimensional input data and limited training examples. GeoRefineNet surpasses the existing state-of-the-art models in the 2D and 3D domains, as demonstrated by its superior performance on numerical and clinical datasets. These findings indicate that GeoRefineNet offers a promising avenue for improving the accuracy and reliability of cephalometric landmark detection fostering further advances in clinical diagnosis and treatment planning. Our code is available at https://github.com/Thanaporn09/GeoRefineNet.git .