| 杨国亮,赖振东,喻丁玲.一种改进UNet++网络用于检测黑色素瘤皮肤病变[J].中国医学影像技术,2020,36(12):1877~1881 |
| 一种改进UNet++网络用于检测黑色素瘤皮肤病变 |
| An improved UNet++ network applied in detection of skin lesions in melanoma |
| 投稿时间：2019-10-15 修订日期：2020-05-28 |
| DOI：10.13929/j.issn.1003-3289.2020.12.025 |
| 中文关键词: 黑色素瘤 机器学习 |
| 英文关键词:melanoma machine learning |
| 基金项目:国家自然科学基金（51365017） |
|
| 摘要点击次数: 1609 |
| 全文下载次数: 731 |
| 中文摘要: |
| 目的 探究基于改进UNet++网络的图像分割算法用于分割黑色素瘤皮肤病变图像的价值。方法 构建引入软注意力门和以Tversky-Focal Loss（TFL）函数为损失函数的UNet++网络优化结构——AT-UNet++网络，并将其在国际皮肤成像协作组织（ISIC）挑战2016和2017训练集中训练。计算训练好的AT-UNet++网络与U-Net网络、UNet++网络的逐像素分割精度（ACC）、DIC相似系数（DIC）和Jaccard相似指数（JAI），对以TFL函数为损失函数的UNet++网络和引入软注意力门的UNet++网络在ISIC挑战2016和2017测试集上进行指标评估；比较ISIC挑战2016与2017竞赛排名前五名的参赛队伍模型与AT-UNet++网络的指标参数。结果 在ISIC挑战2016测试集上，AT-UNet++网络逐的ACC、DIC和JAI较UNet++网络分别提高3.36%、4.15%和3.95%，在2017测试集分别提高2.65%、5.01%及4.39%。结论 AT-UNet++网络的各项评价指标较其他模型均有不同程度提高。 |
| 英文摘要: |
| Objective To explore the value of an improved UNet++ network applied in detection of skin lesions in melanoma. Methods An improved UNet++ network structure—AT-UNet++ network using soft attention gate and Tversky-Focal loss (TFL) function asloss function was proposed. Evaluation metrics was calculated of AT-UNet++ network, U-Net network, UNet++ network, UNet++ network with TFL function and UNet++ network with soft attention gate after training on International Skin Imaging Collaboration (ISIC) Challenge 2016 and 2017 test set, respectively. Then evaluation indexes of the first five competition models in ISIC Challenge 2016 and 2017 Leadership were compared with those of AT-UNet++ network. Results The pixel-wise accuracy (ACC), DICE similarity coefficient (DIC) and Jaccard index (JAI) of AT-UNet++ network based on ISIC Challenge 2016 test set were 3.36%, 4.15% and 3.95% higher than those of UNet++ network, respectively, while ACC, DIC and JAI of AT-UNet++ network based on ISIC Challenge 2017 test set were 2.65%, 5.01% and 4.39% higher than those of UNet++ network,respectively. Conclusion Evaluation indexes of the improved UNet++ network model were improved to different extents compared with those of other models. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
