基于多尺度注意力的皮肤镜图像自动分割算法

针对现有的皮肤镜图像分割算法存在边缘分割时效果较差和对中小目标的识别能力较弱等问题。本文提出了一种基于多尺度注意力融合的分割网络MAU-Net(Multi-scale attention U-Net)。MAU-Net网络是以U-Net网络为基础的分割模型,通过本文设计的多尺度注意力模块(MA),在特征提取时融合不同层次的特征,并将重要的目标特征给与一定的权重,从而使网络能更快和更精准的分割出目标区域。实验结果显示,在ISIC2017数据集上平均交并比(MIOU)、精确度(PRE)和kappa值分别为83.61%、93.58%和81.70%,性能比U-Net分别提高了5.27%、2.01%和6.83%；并在ISIC2017挑战赛数据集上进行了消融实验,实验结果验证了MA模型有助于网络性能的提升。本文提出的MAU-Net网络在皮肤病变分割任务中表现优异,同时具有良好的泛化性能。

Multi-scale Attention for Dermoscopic Image Automatic Segmentation

The existing dermatoscope image segmentation algorithms have some problems, such as poor edge segmentation effect and weak recognition ability for small and medium targets. A segmentation network MAU-Net (multi-scale attention U-Net) based on multi-scale attention fusion is proposed. MAU-Net is a segmentation model based on the U-Net network. Through the design of multi-scale attention model (MA), different levels of features are fused in feature extraction, and important target features are given a weight, so that the network can segment the target area faster and more accurately. The experimental results show that the MIoU(mean intersection over union) precision and kappa values are 83.61%, 93.58% and 81.70% respectively, which are 5.27%, 2.01% and 6.83% higher than those of U-Net. The ablation experiment on ISIC 2017 challenge datasets verifies that MA model is helpful to improve the performance of the network. The MAU-Net network proposed in this paper performs well in the task of skin lesion segmentation and has good generalization performance.