ABSTRACT The annotation of cephalometric landmarks plays a critical role in craniofacial diagnosis and treatment. Compared to conventional manual methods, deep learning‐based automated approaches significantly reduce both time requirements and labour costs. Most current deep learning models primarily rely on convolutional neural networks, but these models exhibit limitations in capturing long‐range dependencies between pixels. Transformer‐based models can effectively address this issue; however, they exhibit poor inductive bias when applied to small‐scale image datasets and are computationally expensive to train on high‐resolution images. In this paper, we propose a novel feature extraction module that reduces the quadratic complexity of computing global attention while enhancing the diversity of global features. Moreover, we extend the range of the original heatmap values and generate multiple outputs for each landmark position prediction. We integrate these components into a simple U‐shaped model, and it achieves competitive detection accuracy without using any pretrained or additional processes compared to several recent methods. In addition, our experiments reveal that the Gaussian kernel size is a critical factor affecting model performance, a parameter that has not been extensively explored in the existing literature.
Xing et al. (2026) studied this question.