New Deep Neural Nets for Fine-Grained Diabetic Retinopathy Recognition on Hybrid Color Space
New Deep Neural Nets for Fine-Grained Diabetic Retinopathy Recognition on Hybrid Color Space
复制标题
DOI：
10.1109/ism.2016.0049
复制
发表时间：
2016-12
期刊：
影响因子：
--
通讯作者：
Holly H. Vo;Abhishek Verma
中科院分区：
文献类型：
--
作者：
Holly H. Vo;Abhishek Verma
糖尿病视网膜病变（DR）的自动识别可以帮助DR携带者在早期接受治疗，避免视力丧失的风险。在本文中，我们强调了多个过滤器大小在学习细粒度判别特征中的作用，并提出了：（i）两个深度卷积神经网络-组合核与多个损失网络（CKML Net）和VGGNet与额外核（VNXK），这是在DR任务背景下对GoogLeNet和VGGNet的改进。从现有的研究学习，（ii）我们提出了一个混合颜色空间，LGI，DR识别通过建议的网络。(iii)迁移学习被应用于解决不平衡数据集的挑战。使用两个大挑战视网膜数据集：EyePACS和Messidor来评估所提出的新网络和颜色空间的有效性。我们的实验结果表明：（iv）CKML Net在使用LGI颜色空间的两个数据集上都优于GoogLeNet，VNXK优于VGGNet。此外，所提出的方法改进了Messidor数据集的其他最新结果，用于可验证/不可验证筛选。
Automatic diabetes retinopathy (DR) recognition can help DR carriers to receive treatment in early stages and avoid the risk of vision loss. In this paper, we emphasize the role of multiple filter sizes in learning fine-grained discriminant features and propose: (i) two deep convolutional neural networks - Combined Kernels with Multiple Losses Network (CKML Net) and VGGNet with Extra Kernel (VNXK), which are an improvement upon GoogLeNet and VGGNet in context of DR tasks. Learning from existing research, (ii) we propose a hybrid color space, LGI, for DR recognition via proposed nets. (iii) Transfer learning is applied to solve the challenge of imbalanced dataset. The effectiveness of proposed new nets and color space is evaluated using two grand challenge retina datasets: EyePACS and Messidor. Our experimental results show: (iv) CKML Net improves upon GoogLeNet and VNXK improves upon VGGNet on both datasets using the LGI color space. Additionally, proposed methodology improves upon other state of the art results on Messidor dataset for referable/non-referable screening.