EIU-Net: Enhanced feature extraction and improved skip connections in U-Net for skin lesion segmentation

Skin lesion segmentation is a computer-aided diagnosis method for quantitative analysis of melanoma that can improve efficiency and accuracy. Although many methods based on U-Net have achieved tremendous success, they still cannot handle challenging tasks well due to weak feature extraction. In response to skin lesion segmentation, a novel method called EIU-Net is proposed to tackle the challenging task. To capture the local and global contextual information, we employ inverted residual blocks and an efficient pyramid squeeze attention (EPSA) block as the main encoders at different stages, while atrous spatial pyramid pooling (ASPP) is utilized after the last encoder and the soft-pool method is introduced for downsampling. Also, we propose a novel method named multi-layer fusion (MLF) module to effectively fuse the feature distributions and capture significant boundary information of skin lesions in different encoders to improve the performance of the network. Furthermore, a reshaped decoders fusion module is used to obtain multi-scale information by fusing feature maps of different decoders to improve the final results of skin lesion segmentation. To validate the performance of our proposed network, we compare it with other methods on four public datasets, including the ISIC 2016, ISIC 2017, ISIC 2018, and PH2 datasets. And the main metric Dice scores achieved by our proposed EIU-Net are 0.919, 0.855, 0.902, and 0.916 on the four datasets, respectively, outperforming other methods. Ablation experiments also demonstrate the effectiveness of the main modules in our proposed network. Our code is available at https://github.com/AwebNoob/EIU-Net.