图像分割是计算机辅助阅片的基础, 伤口图像分割的准确率直接影响伤口分析的结果. 传统方法进行伤口分割步骤繁琐, 准确率低. 目前已有少部分人利用深度学习进行伤口图像分割, 但是他们都是基于小型数据集, 难以发挥深度神经网络的优势, 准确率难以进一步提高. 充分发挥深度学习在图像分割领域的优势需要大型数据集, 目前还没有关于伤口图像的大型公共数据集, 而制作大型伤口图像数据集需要人工标记, 耗费大量时间和精力. 本文提出基于迁移学习的伤口图像分割方法, 首先利用大型公共数据集训练ResNet50网络作为特征提取器, 再利用该特征提取器连接上两个并行的注意力机制后在利用小型伤口图像数据集进行再训练. 实验表明本方法的分割结果在平均交并比上有较大提高, 在某种程度上解决了缺乏大型伤口图像数据集而导致伤口图像分割准确率低的问题.
Image segmentation is the basis of computer-aided film reading, and the accuracy of wound image segmentation directly affects the results of wound analysis. However, the traditional method of wound image segmentation has cumbersome steps and low accuracy. At present, a few studies have applied deep learning to wound image segmentation, but they are all based on small data sets and can hardly give full play to the advantages of deep neural networks and further improve accuracy. Maximizing the advantages of deep learning in the field of image segmentation requires large data sets, but there is no large public data set on wound images as establishing large wound image data sets requires manual labeling, which consumes a lot of time and energy. In this study, a wound image segmentation method based on transfer learning is proposed. Specifically, the ResNet50 network is trained with a large public data set as a feature extractor, and then the feature extractor is connected with two parallel attention mechanisms for retraining with a small wound image data set. Experiments show that the segmentation results of this method are greatly improved in the average intersection over union (IoU), and this method solves the problem of low accuracy in wound image segmentation due to the lack of large wound image data sets to some extent.