[1]邓 辉,张 洁.基于改进的 ResNet50 网络的黑色素瘤分类方法[J].计算机技术与发展,2023,33(02):64-70.[doi:10. 3969 / j. issn. 1673-629X. 2023. 02. 010]
　DENG Hui,ZHANG Jie.Research on Skin Cancer Classification Method Based on Improved ResNet50[J].,2023,33(02):64-70.[doi:10. 3969 / j. issn. 1673-629X. 2023. 02. 010]
点击复制

基于改进的 ResNet50 网络的黑色素瘤分类方法()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
33
期数:
2023年02期
页码:
64-70
栏目:
媒体计算
出版日期:
2023-02-10

文章信息/Info

Title:
Research on Skin Cancer Classification Method Based on Improved ResNet50
文章编号:
1673-629X(2023)02-0064-07
作者:
邓 辉; 张 洁
南京邮电大学 计算机学院,江苏 南京 210023
Author(s):
DENG Hui; ZHANG Jie
School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210023,China
关键词:
黑色素瘤; ResNet50; 注意力机制; GELU; 数据增强; 迁移学习
Keywords:
melanoma; ResNet50; attention mechanism; GELU; data augmentation; transfer learning
分类号:
TP391
DOI:
10. 3969 / j. issn. 1673-629X. 2023. 02. 010
摘要:
黑色素瘤的早期诊断对提高患者的五年生存率至关重要。 针对临床上使用皮肤镜检查黑色素瘤费时、费力的问题,提出一种基于迁移学习和改进的 ResNet50 模型的黑色素瘤分类模型 MC-Net( Melanoma Classification-Net) 。 首先,为了降低数据集样本分布不均匀以及毛发遮挡信息带来的影响,进行数据增强;对 ResNet50 的输入主干重新进行了设计,用常规卷积和深度可分离卷积的组合代替原来的 7×7 大卷积核;对残差块进行了优化,使用 GELU 函数替代 ReLU 函数,并将特征相加后的激活层移到残差块内部,同时去除了部分 BN 层和激活层;向网络中添加 CA 注意力机制,使得网络更加关注目标的关键信息,从而抑制其他无用信息;结合迁移学习,利用在 ImageNet 上预训练权重初始化网络并在数据集上进行微调训练,得到最终黑色素瘤分类模型。 实验结果表明,所改进的结构对模型性能提升具有显著效果,MC-Net 模型在准确率和 F1 指标上达到 94. 87% 和 95. 01% ,经过迁移学习得到的最终分类模型在数据集上进行测试,获得了 95. 4% 的准确率和 95. 6% 的 F1 值,均优于其他网络。
Abstract:
Early diagnosis of melanoma is critical to improving a patient’s five-year survival rate. Aiming at the time-consuming andlabor - intensive problem of clinical use of dermoscopy for melanoma, a melanoma classification model MC - Net ( MelanomaClassification-Net) based on transfer learning and improved ResNet50 model is proposed. Firstly, in order to reduce the impact ofuneven distribution of samples in the dataset and the influence of hair occlusion information,data enhancement is performed. The inputbackbone of ResNet50 is redesigned,and the combination of conventional convolution and depthwise separable convolution is used toreplace the original 7 × 7 large convolution kernels. The residual block is optimized, the GELU function is used to replace the ReLUfunction,the activation layer after feature addition is moved to the inside of the residual block,and some BN layers and activation layersare removed. The CA attention mechanism is added to make the network pay more attention to the key information of the target,therebysuppressing other useless information. Combined with transfer learning,the network is initialized with pre-trained weights on ImageNetand fine- tuned on the dataset to obtain the final melanoma classification model. The experimental results show that the improvedstructure has a significant effect on the performance improvement of the model. The MC-Net model achieves 94. 87% and 95. 01% inthe accuracy and F1 indicators. The final classification model obtained by transfer learning is tested on the data set,and the accuracy of95. 4% and F1 value of 95. 6% are obtained, both superior to other networks.
更新日期/Last Update: 2023-02-10