Children's dental panoramic radiograph image segmentation via dual aggregation transformer UNet

Medical image processing is a critical field in computer vision, aiding in diagnostics and advancing medical research. Specifically, the diagnosis of children's dental conditions often involves the use of panoramic radiographs, which necessitate the segmentation of teeth for accurate diagnosis. However, children's teeth present more complexities than adult teeth due to limited datasets and the substantial changes in size and shape with age, along with the transition period between deciduous and permanent teeth, making segmentation challenging.

This paper aims to introduce a medical image segmentation model specifically designed for children's teeth datasets, capable of efficiently and accurately segmenting children's teeth.

We designed a novel network structure called dual aggregation transformer UNet (DATUnet). The transformer layers aggregate features across spatial and channel dimensions, applying spatial and channel self-attention mechanisms alternately in transformer blocks. This enhances the ability to extract variable shapes and sizes from children's teeth and address the issues like indistinct images and blurriness. Additionally, we developed a multi-scale fusion decoder structure that adaptively fuses shallow and deep information processed through the transformer blocks.

Our model, evaluated on a private dataset of children's dental panoramic radiographs, achieved outstanding performance with a pixel accuracy (PA) of 98.02%, Dice coefficient of 93.20%, and Hausdorff distance (HD) of 4.21.

Compared to other methods, our proposed model achieves superior results, reaching the state-of-the-art (SOTA) in this field.