高度不平衡肠道图像数据集均衡策略

医学图像分类天然受到数据集高度不平衡问题影响。文章首先阐述医学图像的数据不平衡问题及其对人工智能技术分类性能产生的影响。随后研究现有对数据不平衡问题提供类间数据平衡分布的方法,讨论肠道图像中遇到的暗区以及病灶区域旋转等特殊挑战,根据肠道图像的特点,设计一种适用于肠道图像数据集的类间数据平衡方法。核心思路包括利用数量巨大的健康图像为基底,线性组合病灶图像,在选取图像时去除暗区部分较大的健康图像以降低暗区对分类性能的影响,同时通过病灶图像的旋转和平移模拟临床中肠道病灶的旋转特性和位置不确定的特点。通过这种简洁的方法,能够快速地进行高度不平衡的肠道图像数据集平衡。基于肠道图像疾病数据集进行实验,通过肠道溃疡和健康二分类仿真验证文章方法的有效性。

Medical image classification is naturally affected by the high imbalance of data sets.This paper first describes the data imbalance of medical images and its impact on the classification performance of artificial intelligence technology.Then,the existing methods to provide inter-class data balance distribution for data imbalance are studied,and the special challenges such as dark area and focus region rotation in intestine image are discussed.According to the characteristics of intestine image,an inter-class data balancing method suitable for intestine image dataset is designed.The core idea includes using a large number of healthy images as the base to linearly combine the focus images,and removing some large health images in the dark area when selecting the image to reduce the influence of the dark area on the classification performance.At the same time,the rotation and translation of the focus image are used to simulate the rotation characteristics and location uncertainty of intestinal focus in clinic.Through this concise method,the highly unbalanced intestinal focus image data set can be balanced quickly.Experiments are carried out based on the intestine image disease data set,and the effectiveness of the proposed method is verified by intestinal ulcer and health binary classification simulation.