融合多维注意力机制CNN皮肤肿瘤图像分割提取

Combined Multidimensional Attention Mechanism Convolutional Neural Network in Skin Tumor Image Segmentation

摘要:

针对卷积神经网络(CNN)在医学图像分割时,受皮肤病损图像多样性、分割目标位置、形状及尺度变化等因素影响,提出了一种基于传统卷积神经网络综合注意力模块图像分割算法。首先利用U-Net主干网络的优势,其目的让图像特征提取更完善；其次,由空间、通道、尺度构成的综合注意力机制对目标病灶区域进行检测识别,利用通道级联把来自编码器中低级图像特征和解码器中高级图像特征注意力结合起来进行权值自适应融合,提升了网络对样本病灶区的关注度和辨识力,突出强调最相关的特征通道和多尺度间最显著的特征图。通过对ISIC2018数据集及医院整形外科提供患者不同类型的皮肤肿瘤图像进行分割测试,并将注意力模块随机组合形成的不同算法进行指标评价比对,所提出算法的平均分割精度可达92.89%。实验结果表明,所提出算法是有效可行的,在多维度下分割处理带复杂背景的皮肤病灶图像时有更高的鲁棒性。

Abstract:

In medical image segmentation, convolutional neural network (CNN) is affected by the diversity of skin lesions images, the location, shape and scale changes of segmentation targets, and other factors. A multi-dimensional attention module based on space, channel and scale is proposed to optimize the convolutional neural network image segmentation algorithm. Firstly, using the advantage of U-NET backbone network, its purpose is to make image feature extraction more perfect. Second, composed of multidimensional space, channels, scale attention mechanism identification of target lesion area detection, using the lower-level channel cascade from the encoder image features and the decoder senior image adaptive weighting fusion of attention unifies, enhance the awareness of the network on the sample lesions and discrimination, and highlight the most relevant characteristics of the channel, Emphasize the most salient feature maps between multiple scales. The segmentation test was carried out on ISIC2018 data set and images of different types of skin tumors provided by hospital plastic surgery department, and the index evaluation and comparison of different algorithms formed by random combination of attention modules showed that the average segmentation accuracy of the proposed algorithm could reach 92.89%. Experimental results show that the proposed algorithm is effective and feasible, and has higher robustness in the segmentation of cutaneous lesions under complex background.