Swin Transformer Combined with Convolutional Encoder For Cephalometric Landmarks Detection

In orthodontic treatment, the accurate localization of cephalometric landmarks is very essential in clinical diagnosis, treatment and research. In this paper, we propose a transformer-based model named Swin-CE that combines Swin Transformer encoder and convolutional encoder for cephalometric landmark detection. Convolutional encoder can effectively extract local features while Swin Transformer encoder is expert in capturing global features and learning long-range semantic information interactively. Taking advantage of the decoder and skip connections, we integrate the features generated by the two encoders for the final heatmap regression. In experiments, we evaluate our Swin-CE model on a public data set and it achieves state-of-art performances.