Enhanced Landmark Detection in Oral and Maxillofacial CT and CBCT Scans: A Multistage Global–Local Integration Approach

Cephalometric analysis plays a crucial role in oral clinic, especially in orthodontics and maxillofacial surgery, it is essential for precise and reliable treatment planning and diagnosis. Accurate detection of anatomical landmarks is vital for successful cephalometric analysis. Although significant advancements on deep learning-based landmark detection from medical images have been achieved, especially two-stage coarse-to-fine approaches, the performance remains constrained by the high resolution of images and the consequent limitation of computational resources, particularly with three-dimensional (3D) medical data. In this study, we propose a multi-stage approach to precisely detect landmarks from 3D computed tomography (CT) and cone beam computed tomography (CBCT) images. First, we utilize an alignment network to identify the region of interest and reduce redundant information. Next, we introduce a novel coarse-to-fine detection framework employing 3D U-Net model with heatmap regression as backbones. Specially, we propose a multi-scale global information extraction module (MSGIEM) that combines features from multiple layers with varying resolutions and semantics to capture effective global information. Additionally, a landmark attention module (LAM) is introduced to distill relevant global features for individual landmarks, facilitating their integration with local visual patches to enhance overall detection performance. Extensive experiments on in-house CT and CBCT datasets, as well as a public dataset, demonstrate that the proposed method significantly outperforms state-of-the-art approaches, with a notable reduction on the mean radial error (MRE) by 0.14 mm and 0.04 mm, and an improvement on the successful detection rate (SDR) within 2 mm by 3.98% and 0.23% on in-house CT and CBCT datasets respectively. Furthermore, our discussion on the detection of common landmarks across CT and CBCT datasets underscores its potential clinical applications. The code is available at https: //github.com/HLZ-CODE/Maxillofacial-Landmark.