多尺度特征融合的改进残差网络乳腺癌病理图像分类
Improved Residual Network Classification of Breast Cancer Pathological Images Based on Multi-Scale Feature Fusion
庄建军 1吴晓慧 1景生华 2孟东东1
作者信息
- 1. 南京信息工程大学电子与信息工程学院,南京 210044
- 2. 东部战区总医院放疗科,南京 210002
- 折叠
摘要
现有模型病理特征提取不充分以及开源数据集各类型数量不平衡等问题,使得乳腺癌病理图像的多分类研究仍具挑战性.本研究提出了一种多尺度特征融合的改进残差网络乳腺癌病理图像多分类方法.首先,以ResNet101残差网络作为基础,将CBAM注意力模块插入到每一个残差块中;接着,为了优化特征提取,将横向和纵向的多尺度特征融合集成到残差网络中;最后,引入焦点损失函数以解决数据分配不平衡问题.经BreakHis公开数据集混合放大倍数1 582张病理图像训练验证,所提出的改进残差网络在乳腺癌病理图像八分类上的识别准确率为94.4％,较原始模型提升2.8％,优于大多数已有公开深度学习模型.该模型的提出为女性乳腺癌的筛查诊断和病理分类提供了更为有效的方法.
Abstract
In view of the extremely inadequate extraction of pathological features from existing models and the unbalanced number of various types of open breast cancer data sets,the research on multi-classification of breast cancer pathological images is still challenging.In this paper,an improved residual network multi-classification method of breast cancer pathological images based on multi-scale feature fusion was proposed.Firstly,based on ResNet101 residual network,the CBAM attention module was inserted into each residual block.Next,in order to optimize feature extraction,horizontal and vertical multi-scale feature fusion was integrated into the residual network.Finally,the focus loss function was introduced to solve the problem of unbalanced data distribution.Validated by the training of 1582 pathology images with mixed magnifications on BreakHis public dataset,the proposed improved residual network achieved a recognition accuracy of 94.4％on eight classifications of breast cancer pathology images,which was 2.8％better than the original model and outperforms most of the existing publicly available deep learning models.The proposed model provided a more effective method for screening,diagnosis and pathological classification of female breast cancer.
关键词
乳腺癌病理图像/深度学习/残差网络/注意力机制/多尺度特征融合Key words
pathological images of breast cancer/deep learning/residual network/attention mechanism/multi-scale feature fusion引用本文复制引用
基金项目
国家重点研发计划(2021YFE0105500)
国家自然科学基金(62171228)
江苏高校'青蓝工程'资助()
出版年
2024
NETL
NSTL
万方数据