基于改进的AG-CNN的视网膜OCT图像的黄斑病变识别方法

Macular degeneration identification method of the OCT image with retina based on improved AG-CNN

本文针对目前应用全局图像训练卷积神经网络可能会受到若干无关噪声区域的影响,易导致视网膜OCT图像黄斑病变识别或诊断错误等问题,提出了一种改进的注意力引导四分支卷积神经网络的视网膜OCT图像黄斑病变识别方法。采用改进注意力引导卷积神经网络框架,通过集成全局分支、局部分支和层分割分支构成融合分支,利用注意力热图对重要区域进行掩膜和训练,减少视网膜OCT图像噪声的干扰和黄斑病变识别错误率,通过与VGG16和IDL2种方法在公开数据集上进行了实验验证比较。结果表明,文中方法在视网膜OCT图像数据集上对于识别准确度和识别性能的提升具有显著性的作用。

In view of the current application of global image training convolutional neural network may be affected by a number of irrelevant noise regions,which may easily lead to errors in the recognition or diagnosis of macular disease in retinal OCT images,an improved attention-guided four-branch convolutional neural network method was proposed to identify the macular degeneration in retinal OCT images. The attention-guided convolutional neural network framework was used to form fusion branches by integrating global branches,local branches and layer segmentation branches,and attention heat maps were used to mask and train important areas,which reduced the interference of retinal OCT image noise and macular disease recognition error rate,the proposed was compared with the two methods of VGG16 and IDL on the public data set. The results showed that the proposed method had a significant effect on the improvement of recognition accuracy and recognition performance on the retina OCT image data set.