LTB-Net: A Lightweight Transformer-based Burn Depth Segmentation Network

Abstract:
Burn is a common traumatic injury, and early diagnosis of burn depth is crucial for reducing treatment costs and improving patient survival rates. However, burn depth segmentation remains a challenging task because of the diversity in shape and size, and also the striking similarity in color between burn wounds of different depths. Current clinical burn depth diagnosis heavily relies on subjective visual inspection by clinicians, leading to low accuracy. Transformers, with their advantage in modelling long-range dependencies, have recently found applications in medical image segmentation. However, the lack of local information in transformers may result in inaccurate burn depth recognition. Moreover, increasingly complex structures lead to lower efficiency, making direct application challenging. In this paper, we apply the transformer to the field of burn segmentation for the first time, proposing a lightweight transformer-based network for burn depth segmentation, named LTB-Net. It includes a transformer-based encoder and an improved MLP-based decoder, significantly reducing the number of parameters. Additionally, to enhance local information acquisition, we introduce a local information enhancement module (LIEM). Finally, we introduce a background information filtering module (BIFM) to prevent information loss during down-sampling in small burn wound regions and further enhance information extraction in LIEM. Extensive experiments on our burn dataset demonstrate the superiority of our model compared to other models.