【摘要】 深度卷积神经网络是图像分类领域的主流算法之一，但是其训练需要大量标注数据，在阿尔茨海默病医学影像等小数据集上易出现过拟合现象。数据增强算法可以用于扩充数据量，其中CutMix数据增强算法近来被广泛应用，但是现有方法生成的增强图像往往忽略原始图像显著区域，且增强图像的标签设计考虑的因素较为单一。针对这些问题，提出Fit CutMix数据增强算法。该算法一是利用基于显著性极值迁移的区域替换策略生成增强样本，集中源样本与目标样本中显著性高的区域；二是综合源样本与目标样本的面积和显著性信息赋予增强样本标签，为卷积神经网络提供有效的监督信息。实验结果表明，将Fit CutMix数据增强算法用于ResNet50对阿尔茨海默病进行诊断时，准确率达96.6%，比直接使用ResNet50提高了约7%，且比应用现有数据增强算法至少提高3%，可见Fit CutMix数据增强算法可以有效提高深度卷积神经网络对医学影像识别的准确率。更多还原

【Abstract】 Deep convolutional neural network is one of the mainstream algorithms in the field of image classification, but its training requires a large number of labeled data, which leads to over fitting on small datasets such as Alzheimer’s medical images. Data augmentation can increase the amount of training data, and Cut Mix data augmentation algorithm has been widely used recently. However, the augmented images generated by the CutMix series methods often ignore the significant area of the original image, and the design of the label of the augmented image takes only single factor into consideration. In order to solve these problems, the Fit Cut Mix data augmentation algorithm was proposed. Firstly, the region replacement strategy based on the transfer of saliency extreme value was used to generate augmented samples, so as to concentrate the regions with high saliency value in the source samples and target samples. Secondly, the area and saliency information of the source samples and the target samples were combined to assign the augmented sample label, which provided effective supervision information for the convolutional neural network. The experimental results showed that when Fit CutMix was used in Res Net50 to diagnose Alzheimer’s disease, the accuracy was 96.6%, which was about 7% higher than that of directly using Res Net50, and at least 3% higher than that of applying existing methods. Therefore, the Fit CutMix data augmentation algorithm can effectively improve the recognition accuracy of deep convolutional neural network for medical images.更多还原