Automatic Skin Lesion Segmentation Based on Higher-Order Spatial Interaction Model

Dermoscopic image segmentation is a key step in computer-aided diagnosis of skin lesions. The current medical image segmentation model mainstream uses standard convolution and Transformers as the most important components of the model. However, standard convolution is unable to handle the problems of remote information interaction and long distance spatial dependence. Transformers is a hindrance to its application when dealing with medical clinical data, the insufficient amount of data and the large memory and time required for computation are hindering factors. Recently, HorNet, a high-order spatial interaction model that performs well in natural scenarios, has drawn our attention because the high-order interaction model has the common advantages of both convolution and Transformers. In this paper, we propose a new system that uses the HorNet model, which performs well in natural scenes, for medical image segmentation (skin lesion segmentation). To the best of our knowledge, we are the first to use a higher-order interaction model for medical image segmentation. We validate our system on three public datasets and do external validation on our own clinical dataset. The experimental results show that our system outperforms other current medical image segmentation models in several metrics. We think this is due to the higher spatial interaction capability and larger perceptual domain of gnconv in HorNet, which can fully and comprehensively capture the information of the overall medical image, which is crucial for medical image processing.