Abstract
Accurate skin lesion segmentation is of great significance for improving the quantitative analysis of skin cancer. However, due to the irregular shapes and ambiguous boundaries of lesions, automatic segmentation is still very difficult. Although the emergence of CNN and Transformer hybrid models recently complements their respective shortcomings, the use of transformer results in heavy parameters, and the complex structure requires a lot of data and resources for training. In contrast, Multi-Layer Perceptron (MLP) is a cost-effective alternative to complex self-attention. To this end, we propose an MLP-based model using matrix decomposition and rolling tensors for skin lesion segmentation. Specifically, we can dynamically calculate the correlation matrix according to the size of the input image and use it as the weight to guide the segmentation. The way of rolling tensors can fully mix the feature information and sequentially extract the information under different receptive fields. We also employ a two-layer decoding structure using matrix decomposition to fuse the feature extracted in the parallel MLP encoders. Extensive experiments are conducted on the public skin disease datasets ISIC2016, ISIC2017, ISIC2018 and PH2. After comparison, we found that although the addition of Transformer makes up for the lack of global information on CNN, the complex structure of Transformer itself will lead to huge parameters and is difficult to train. The experimental results show that we achieved higher results with fewer parameters compared to other state-of-the-art models which mean that the self-attention mechanism of Transformer is not irreplaceable. Our code is released at https://github.com/Lingglesymphony/RMMLP.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have