非局部卷积残差学习模型的病理图像分类方法

Pathological image classification based on non-local convolution residual learning model

针对组织病理学图像癌细胞分布随机性强、分布广泛的特点,且常见的卷积神经网络(convolutional neural network,CNN)难以直接获取长范围依赖关系的问题,本文采用基于迁移学习的ResNet与优化后的Non-local Net相结合的非局部卷积残差模型,融合了位置和通道特征,在图像中提取全局信息选择有价值的区域进行分类。本文采用BreakHis数据集进行实验,针对数据集良恶性样本分布极不均衡的问题,采用Random-SMOTE算法平衡良恶性样本,强化模型学习少数类别特征的能力。在不区分倍数的数据集上,本文提出方法的PR、RE、SP和ACC分别达到93.28%、98.71%、98.67%和98.70%;在已知倍数的数据集上,上述指标也更高。与乳腺癌组织病理学分类中常用的算法相比,本文提出的方法具有更好的性能。

Aiming at the characteristics of strong randomness and wide distribution of cancer cells in histopathological images,and it is difficult for common convolutional neural network(CNN)to directly obtain long-range dependencies,this thesis adopts ResNet based on transfer learning and optimized Non-local Net.The combined non-local convolutional residual model fuses location and channel features to extract global information in images for selecting valuable regions for classification.In this thesis,the BreakHis dataset is used for experiments.In view of the extremely unbalanced distribution of benign and malignant samples in the dataset,the Random-SMOTE algorithm is used to balance benign and malignant samples to strengthen the model's ability to learn minority class features.On the datasets that do not distinguish multiples,the PR,RE,SP and ACC of the proposed method reach 93.28%,98.71%,98.67%and 98.70%,respectively;on the datasets with known multiples,the above indicators are also higher.Compared with the commonly used algorithms in breast cancer histopathological classification,the method proposed in this thesis has better performance.