融合动态数据增强与MACA的皮肤病分类方法  

Skin Lesion Classification Method Combining Dynamic Data Augmentation and MACA

作　　者：李志明 蒋舒颖 向凡 张俊然[1] LI Zhi-ming;JIANG Shu-ying;XIANG Fan;ZHANG Jun-ran(College of Electrical Engineering,Sichuan University,Chengdu Sichuan 610065,China)

机构地区：[1]四川大学电气工程学院,四川成都610065

出　　处：《计算机仿真》2024年第12期264-269,437,共7页Computer Simulation

基　　金：国家自然科学基金数学天元基金重点专项(12126606);德阳科技(揭榜)项目(2021JBJZ007);智能电网四川省重点实验室应急重点项目(.020IEPG-KL-20YJ01);华西医院135交叉学科创新项目(ZYJC21041);四川大学“医学+信息”中心融合创新项目(YGJC012)。

摘　　要：针对皮肤病数据集类别不平衡和类间样本相似度高导致不易区分的问题,提出了一种融合动态数据增强与MACA-MobileNet-V2的皮肤病分类方法。首先,设计了一种动态数据增强方法,根据模型表现情况动态地进行数据增强,更充分利用数据集中各类别已有信息,以减小类别不平衡带来的影响;其次,在坐标注意力的基础上提出了多尺度聚合坐标注意力,结合注意力搭建的网络(MACA-MobileNet-V2)可以捕获多尺度远程依赖信息;最后,引入Focal Loss函数和部分参数迁移学习,进一步优化模型。所提方法在皮肤病公开数据集八分类任务中平均准确率、精确度、召回率分别提高5.6%,6%,5%,减小了皮肤病数据集不平衡与类间样本相似度高给分类带来的影响。The skin lesion dataset has the characteristics of class imbalance and high similarity samples between classes,so it is difficult to classify skin lesions.To deal with these problems,we propose a skin lesion classification method combining dynamic data augmentation and MACA-MobileNet-V2.Firstly,a dynamic data augmentation method is proposed to maximize the use of the existing information of each class in the dataset and reduce the impact of class imbalance,it dynamically augments data according to the performance of the model.Secondly,a multi-scale aggregated coordinate attention is proposed based on coordinate attention,and we use it to build MACA-MobileNet-V2 that can capture multi-scale long-range dependency information of features.Finally,the Focal Loss function and partial parameter transfer learning are introduced to optimize the model.The method we propose increases the average accuracy,precision,and recall rate by 5.6%,6%,and 5%in the eight-class task of the public skin lesion dataset,the impact of imbalance and high similarity between classes is reduced.

关 键 词：动态数据增强 类不平衡 多尺度聚合坐标注意力 皮肤病分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] R814[自动化与计算机技术—计算机科学与技术]