基于自适应多分辨率特征学习的CNV分型网络  

Adaptive multi-resolution feature learning network for CNV classification

作　　者：许传臻 袭肖明 李维翠 孙仪 杨璐 XU Chuanzhen;XI Xiaoming;LI Weicui;SUN Yi;YANG Lu(School of Computer Science and Technology,Shandong Jianzhu University,Jinan 250101,Shandong,China;Shandong Institute of Scientific and Technical Information,Jinan 250101,Shandong,China;School of Architectural Urban Planning,Shandong Jianzhu University,Jinan 250101,Shandong,China)

机构地区：[1]山东建筑大学计算机科学与技术学院,山东济南250101 [2]山东省科学技术情报研究院,山东济南250101 [3]山东建筑大学建筑城规学院,山东济南250101

出　　处：《山东大学学报（工学版）》2022年第4期69-75,共7页Journal of Shandong University（Engineering Science）

基　　金：山东省自然科学基金重大基础研究资助项目(ZR2021ZD15);山东省高等学校青创科技支撑计划创新团队(2021KJ036)。

摘　　要：为解决不同脉络膜新生血管(choroidal neovascularization,CNV)类型间较小区分性带来的分型难度和光学相干断层扫描(optical coherence tomography,OCT)图像中噪声对分型精度的影响,提出自适应多分辨率特征学习的CNV分型方法,其包含多分辨率特征学习和自适应特征选择模块。在多分辨率特征学习模块中,融合具有不同类型CNV细节信息的底层特征和具有语义信息的高层特征,同时引入渐进式的训练方式增强特征表示能力。在自适应特征选择模块中,通过引入注意力机制,对最后分型起关键作用的特征进行增强,进一步提升特征的区分性。在自建的CNV数据集上进行试验,试验结果表明,评价指标上的测试评分分别为91.3%、86.6%、89.2%和90.6%。提出的自适应多分辨率特征学习的CNV分型方法优于现有的其他分类方法。The large similarity between different choroidal neovascularization(CNV) types increased the difficulty of classification and the noise in optical coherence tomography(OCT) images affected the classification accuracy.To address the issues,an adaptive multi-resolution feature learning CNV classification network(AMFL-net) was proposed,which consisted of multi-resolution feature learning and adaptive feature selection modules.In the multi-resolution feature learning module,the low-level vision features and the high-level semantic features were fused.The progressive training method was introduced to enhance the feature representation ability.In the adaptive feature selection module,the attention mechanism was introduced to enhance the features that played a key role in the final classification,and further improved the discrimination ability of features.Experimental results showed that the test scores on the evaluation indicators were 91.3%,86.6%,89.2% and 90.6% respectively,and the proposed AMFL-net outperformed traditional classification methods.

关 键 词：CNV分型 自适应多分辨率特征学习 光学相干断层扫描图像 自适应特征选择 多分辨率特征学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]