基于EfficientNetV2和迁移学习的乳腺癌病理图像分类  

Classification of Breast Cancer Pathological Images Based on EfficientNetV2 and Transfer Learning

作　　者：杨传德 李格璇 李海军[1] YANG Chuande;LI Gexuan;LI Haijun(Dezhou University,Dezhou 253023,China)

机构地区：[1]德州学院,山东德州253023

出　　处：《现代信息科技》2023年第10期136-139,共4页Modern Information Technology

基　　金：德州市企业研发计划项目(2022dzkj094);山东省大学生创新训练项目(S202210448029)。

摘　　要：针对人工识别乳腺癌肿瘤存在无法避免人为因素的问题,提出一种基于EfficientNetV2和迁移学习的乳腺癌病理图像分类算法。首先,由于现有公开数据集BreaKHis样本数据量过小,通过数据预处理实现数据增强,从而使模型学习到更多鲁棒性的特征。然后将数据集在微调后的EfficientNetV2网络模型上进行迁移学习训练,同时采用Nadam进行梯度下降优化,进而实现乳腺癌病理图像的自动分类。实验结果表明,相较于其他深度学习模型,该模型以更少的参数与更高的准确率很好地以底层视觉特征映射了高层语义,有效提升了临床医学诊断的效率。Aiming at the problem that artificial recognition of breast cancer tumors cannot avoid human factors,a classification algorithm of breast cancer pathological images based on EfficientNetV2 and transfer learning is proposed.Firstly,due to the small sample size of the existing public dataset BreakKHis,data augmentation is achieved through data preprocessing,thereby enabling the model to learn more robust features.Then,transfer learning training is carried out on the database on thefinely tuned EfficientNetV2 network model,and the gradient descent optimization is carried out with Nadam,so as to realize the automatic classification of breast cancer pathological images.The experimental results show that compared to other deep learning models,this model effectively maps high-level semantics with lower level visual features with fewer parameters and higher accuracy,effectively improving the efficiency of clinical medical diagnosis.

关 键 词：EfficientNetV2模型 迁移学习 NADAM 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] R737.9[自动化与计算机技术—计算机科学与技术]