MCST-Net: A Multi-Cross-Spatial Attention U-Net with Transformer Block for Skin Lesion Segmentation

Abstract

Applying deep learning to skin lesion image segmentation has grown in popularity over the past few years. It is simpler to understand the injuries and administer the proper care using segmented images. In this research, we proposed a new model, by building a multi-cross-attention block, combining it with the transformer block and Unet model. In particular, our multi-cross-spatial attention block is highly effective when it comes to extracting features from multiple layers, helping to increase efficiency in the image upsample process. The transformer block that is widely used in natural language processing (NLP) is incorporated into the model, which helps to improve performance with processing the highest features. Additionally, our model has the benefit of having a small number of parameters in addition to providing great performance on the two well-known datasets, ISIC 2018 and PH2.