MAUNet:用于皮肤病变分割的轻量级模型

针对当前深度学习分割算法参数数量多和计算复杂度高的问题，提出了一种融合多种注意力机制的轻量级模型MAUNet用于皮肤病变分割。该模型在UNet网络基础上融合深度可分离卷积和门控注意力机制模块，用于提取全局和局部特征信息；融入外部注意力机制模块来增强样本间的联系；利用空间和通道注意力模块分别提取通道和空间特征。以ISIC2017皮肤病公开数据集作为数据源，改进的UNet模型实现特征提取与分类。与基线模型UNet相比，平均交并比和Dice相似性系数分别提高了2.18%和1.28%，同时参数量和计算复杂度仅为基线模型的2.1%和0.58%。实验结果表明该模型在参数数量平衡性、计算复杂度和分割检测性能上均达到了较好的水平。

MAUNet: a lightweight model for skin lesion segmentation

Abstract: The current depth learning segmentation algorithm has the problems of numerous parameters and high computational complexity. Therefore, a lightweight algorithm (MAUNet) which combines UNet and multiple attention mechanisms is proposed for skin lesion segmentation. The model integrates depth-wise separable convolution and gated attention mechanism modules on the basis of UNet to extract global and local feature information, adopts the external attention mechanism module to enhance the connection between samples, and uses the spatial and channel attention mechanism modules to extract channel and spatial features. The MAUNet model realizes feature extraction and classification on ISIC2017 skin disease public data set. Compared with the baseline model (UNet), the proposed model increases mIoU and DSC by 2.18% and 1.28% respectively, while reducing the number of parameters and computational complexity which were only 2.1% and 0.58% of the baseline model. The experimental results show that the model can balance the number of parameters, lower the computational complexity and perform well in segmentation and detection.