基于EfficientNet模型的糖尿病视网膜图像分类方法研究

Diabetic Retinal Image Classification Method Based on EfficientNet Model

目的:糖尿病性视网膜病变(diabetic retinopathy, DR)是导致糖尿病患者失明的主要原因。临床上主要通过医生对DR图像进行分析来判断患者是否需要进行治疗。但基于人工的分类性能受到医生主观因素及图像不同等级之间差异微小的影响,会出现误诊漏诊等情况。由此提出融合空间注意力机制的EfficientNet网络模型(A-U EfficientNet)实现DR图像自动准确五分类。方法:空间注意力机制使用FRelu作为激活函数实现像素级的空间信息建模能力。该机制将生成的注意力图与EfficientNet网络的特征进行元素点乘,对患病区域的特征进行增强,使模型自动关注微动脉瘤等相关区域特征并提供分类依据;针对交叉熵函数(cross entropy, CE)未能考虑到样本不均衡问题,引入Focal Loss损失函数,减少易分类样本的权重,使得模型在训练时更专注于难分类的样本。结果:在Kaggle数据集上进行实验后证明,A-U EfficientNet在测试集上的ACC、SP和SE分别为95.2%、97.9%和93.7%,比原始网络分别提高了2.8%、2.6%和3.7%。结论:融合空间注意力机制的EfficientNet网络模型能够有效提高DR分类准确性并向医生提供病灶位置,在临床上避免发生误诊漏诊的情况,防止因治疗不及时而造成严重的视力损伤。

Objective: Diabetic Retinopathy(DR) is the main cause of blindness in diabetic patients. Clinically, doctors mainly analyze DR images to judge whether patients need treatment. However, the classification performance based on manual is affected by the subjective factors of doctors and the slight difference between different levels of images, which will lead to misdiagnosis and missed diagnosis. Therefore, an efficient net model(A-U Efficientnet) integrating spatial attention mechanism is proposed to realize automatic and accurate five classification of DR images. Method: The spatial attention mechanism uses FRelu as the activation function to realize the spatial information modeling ability at the pixel level. The mechanism multiplies the generated attention map with the features of efficientnet network, enhances the features of the diseased area, makes the model automatically pay attention to the relevant regional features such as micro aneurysms and provides classification basis;Aiming at the problem that the cross entropy function(CE) fails to consider the imbalance of samples, the focal loss loss function is introduced to reduce the weight of easy to classify samples, so that the model focuses more on difficult to classify samples during training. Result: The test results showed that the A-U EfficentNet model was more efficient than the original network, and its ACC, SP and SE ratios were 95.2%, 97.9% and 93.7%, respectively, which were 2.8%, 2.6% and 3.7% higher than that of the original network. Conclusion: Efficientnet network model integrating spatial attention mechanism can effectively improve the accuracy of DR classification and provide the location of lesions to doctors, avoid misdiagnosis and missed diagnosis in clinic, and prevent serious visual damage caused by untimely treatment.