基于改进残差网络的黑色素瘤图像分类  

Melanoma image classification based on improved ResNet

作　　者：徐慧 邹俊忠[1] 张见[1] 陈兰岚[1] XU Hui;ZOU Jun-zhong;ZHANG Jian;CHEN Lan-lan(School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237,China)

机构地区：[1]华东理工大学信息科学与工程学院,上海200237

出　　处：《计算机工程与设计》2023年第5期1495-1501,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61976091)。

摘　　要：为提升黑色素瘤图像二分类问题的准确率,针对黑色素瘤图像中存在的有效特征不明显的问题,借鉴特征金字塔思想,提出一种改进的残差网络的黑色素瘤图像分类模型。使用迁移学习,以预训练的ResNet50模型为基础结构,利用改进的注意力机制筛选有效特征,用空洞卷积改进Inception结构并基于该结构构建额外的分支以不同方式提取并融合特征,用加权的方式把分支的特征和ResNet50模型主干提取的特征进行融合。所提模型在ISIC 2017数据集上可以取得87.8%分类准确率,表明了其对解决黑色素瘤图像二分类问题的有效性。To improve the accuracy of melanoma binary classification and address the problem of unapparent effective features in melanoma images,an improved classification model for melanoma images with ResNet networks was proposed.Transfer learning was used and a pre-trained ResNet50 model was taken as the base structure for filtering effective features using an improved attention mechanism.The Inception structure was improved with dilated convolution,and an additional branch was constructed based on this structure to extract and fuse features in different ways.The features of the branch were fused with the features extracted from the backbone of the ResNet50 model in a weighted manner.The proposed model can achieve 87.8%classification accuracy on the ISIC 2017 dataset,demonstrating its effectiveness for solving the melanoma image binary classification problem.

关 键 词：黑色素瘤分类 残差网络 特征分支 特征融合 注意力机制 迁移学习 空洞卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]