Paying attention to the minute details: Supervised keypoint detection on dense, complex point clouds

Keypoints in three-dimensional (3D) objects are crucial for applications such as shape description, registration, and clinical diagnosis. However, current keypoint detection methods for 3D objects often depend on multiview two-dimensional (2D) images or unsupervised learning, largely due to the limited availability of annotated datasets. This limitation restricts their ability to detect anatomically significant keypoints directly on 3D models. To overcome this challenge, we propose a supervised keypoint detection framework designed for dense and complex 3D point clouds. Our framework combines a specialized segmentation network to identify potential keypoint regions, followed by precise keypoint localization using a PointNet encoder with residual modules and a novel “double SoftMax” mechanism. While keypoint detection accuracy benefits from restricting potential regions to smaller areas, this can leave some regions undetected in dense point clouds. To address this, we introduce a “Penalty Dice Loss” into the segmentation network, which effectively reduces undetected regions. Experiments on in-house skull and tibia datasets show mean radial errors of 1.43 mm and 1.54 mm, and success detection rates of 76.00% and 76.40%, respectively, within the clinically accepted 2 mm range. These results are competitive with state-of-the-art 2D cephalometric X-ray landmark detection methods and demonstrate the potential of our framework for clinical applications, such as preoperative planning, intraoperative navigation, and postoperative evaluation.