基于迁移学习的糖尿病视网膜病变眼底图像分类研究

A Study on Image Classification of Diabetic Retinopathy Based on Transfer Learning

摘要: 全世界有4亿多人患有糖尿病，中国的患病人数超过1个亿。而其中1/3的糖尿病患者会出现可能导致视力损伤甚至失明的并发症——糖尿病视网膜病变。因此，检查眼底是每个患有糖尿病的人都需要做的检查，病情越重检查频率越高。而如此多的检查给医疗人员造成很大的工作负担，同时也增加了国家的医疗支出。为了更经济、高效和准确地检查，需要建立计算机自动诊断系统。因此，这项工作成为科研人员重要的研究课题。随着深度学习在图像分类任务中取得的重大进步，利用这一技术建立高效的糖尿病自动诊断系统成为研究热点。然而，由于医疗数据自身的特点(属于小数据集并且数据不平衡)以及神经网络的“黑箱”问题，使得利用深度学习进行疾病分类效果不理想。针对以上问题，本文从以下方面进行研究：提出了一个基于迁移学习和DenseNet121的可解释性分类器。在预先训练的DenseNet121网络上，利用微调迁移学习技术，更改部分网络结构，实现DR自动诊断任务，实验结果为敏感性0.77、特异性0.91、准确率88%。同时对神经网络解释性可视化，采用梯度加权类别激活映射对分类器的分类结果可视化。通过这种方式，可显示输入图像的哪个部分对分类器的分类结果影响更大。通过这种可视化的方式发现，虽然分类器达到了较高的准确率，但分类器并非总是关注那些病灶部位，并且存在对类别少的眼底图像样本的偏见，反映出不平衡样本所训练的模型存在的问题，这让我们更好地了解我们所训练的模型所存在的不足之处。

Abstract: More than 400 million people worldwide suffer from diabetes and more than 100 million in China. A third of diabetics have complications that can lead to visual impairment or even blindness, namely, diabetic retinopathy. Therefore, fundus examination is a necessary check for everyone with diabetes, and the more severe the condition, the higher the frequency of the check. Such a large number of examinations places a heavy workload on medical staff and also increases the country’s medical expenditure. To conduct the examination more economically, efficiently, and accurately, it is necessary to establish a computer-aided diagnostic system. Hence, this work has become an important research topic for scientific researchers. However, due to the characteristics of medical data itself (belonging to small datasets and data imbalance) as well as the “black box” issue of neural networks, the effect of using deep learning for disease classification is not satisfactory. In response to the above problems, this paper conducts research in the following aspects: an interpretable classifier based on transfer learning and DenseNet121 is proposed. By utilizing fine-tuning transfer learning techniques on the pre-trained DenseNet121 network and modifying some of the network structures, the automatic diagnosis task of DR is realized. The experimental results show a sensitivity of 0.77, specificity of 0.91, and accuracy of 88%. At the same time, the interpretability of the neural network is visualized by using gradient-weighted class activation mapping to visualize the classification results of the classifier. Through this method, it can be shown which part of the input image has a greater impact on the classification results of the classifier. By this visualization approach, it is found that although the classifier has achieved a high accuracy rate, it does not always focus on the lesion areas, and there is a bias towards fundus image samples with fewer categories. This reflects the problems existing in models trained with imbalanced samples, allowing us to better understand the shortcomings of the models we have trained.