3D Ear Segmentation Based on Multiscale Point Transformer

Accurate 3D ear point cloud segmentation is crucial for ear-worn device design and Traditional Chinese Medicine (TCM) auricular diagnosis. Traditional methods struggle with disordered point clouds, uneven density, and complex ear geometries, leading to low segmentation accuracy. We propose MSPT_EarSeg, a novel network based on the Point Transformer V3 architecture that integrates serialized attention and multi-scale feature fusion to segment eight auricular acupoint regions. By leveraging Hilbert curve-based serialization and dynamic window attention, MSPT_EarSeg achieves superior performance over PointNet++ and Point Transformer V2 in mean Intersection over Union (mIoU), accuracy, and recall. Experimental results demonstrate its robustness and efficiency, supporting automated ear-worn device design and TCM acupoint diagnosis.