基于特征类内紧凑性的不平衡医学图像分类方法  

Imbalanced medical image classification based on intra⁃class compactness of features

作　　者：孟元 张轶哲 张功萱[1] 宋辉 Meng Yuan;Zhang Yizhe;Zhang Gongxuan;Song Hui(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing,210094,China)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094

出　　处：《南京大学学报（自然科学版）》2023年第4期580-589,共10页Journal of Nanjing University（Natural Science）

基　　金：国家自然科学基金(62272232,62201263);江苏省自然科学基金(BK20220949)。

摘　　要：近些年,基于深度学习的算法和模型在各种图像分析任务中都取得了显著的成功,与常见的自然图像相比,医学图像数据集依然面临高度不平衡的问题,不平衡数据会导致特征空间里的决策边缘倾向样本多的类别,导致分类效果的下降.为了解决该问题,提出一种基于卷积神经网络考虑特征类内紧凑性的不平衡医学图像分类方法(Z-Score Compactness-based Convolutional Neural Network,ZC3NC).首先,从一个卷积神经网络的最后一层卷积层提取训练集样本与测试集样本的特征图,随后引入一个新的Z分数来度量测试集数据的特征图相对训练集每个类在特征空间上的偏离度,偏离度的度量基于类内的紧凑度,其主要关注样本的分布特性,对各类样本数量的不平衡性不敏感.最终,根据计算的偏离度,对测试集的数据进行分类.在DermaMNIST数据集上的实验表明,在不对数据和神经网络模型做任何额外增强的情况下,该方法的平衡准确率比原卷积神经网络模型平均提高11.15%,最多提高14.08%,证明提出的分类方法能有效地提高多种卷积神经网络对不平衡医学图像数据的分类性能.此外,和最先进的不平衡分类方法Under-Bagging KNN相比,该方法的性能平均提升了2.36%.In recent years,algorithms and models based on deep learning have achieved significant success in various image analysis tasks.However,compared to common natural images,medical image datasets often face highly imbalanced problems,which lead to decreased classification performance.Imbalanced data causes decision boundaries in the feature space to tend towards the class with more samples.To solve this problem,this paper proposes an imbalanced medical image classification method based on convolutional neural networks considering intra⁃class compactness of features(Z⁃Score Compactness⁃based Convolutional Neural Network,ZC3NC).First,feature maps of training and testing set samples are extracted from the last convolutional layer of a convolutional neural network.Then,we introduce a new Z⁃score based measure to test the deviation of the testing set data feature maps relative to each class of the training set in the feature space.The measure of deviation is based on intra⁃class compactness,which focuses on the distribution characteristics of the samples and is insensitive to the imbalance of the number of samples in each class.Finally,based on the calculated deviation,we classify the testing set data.Experiments on the DermaMNIST dataset show that without any additional data or neural network model enhancements,the balanced accuracy of the proposed method increases by an average of 11.15%compared to the original convolutional neural network model,with a maximum increase of 14.08%.This verifies that the proposed classification method effectively improves the classification performance of various convolutional neural networks for imbalanced medical image data.Furthermore,compared to the state⁃of⁃the⁃art imbalanced classification method,Under⁃bagging KNN,the average improvement of ZC3NC is 2.36%.

关 键 词：卷积神经网络 类别不平衡 医学图像 特征 分类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]