融合注意力线性特征多样化的DR分级模型

DR grading model of fusing attention linear feature diversification

摘要:

糖尿病视网膜病变(diabetic retinopathy,DR)是目前人类的主要致盲疾病之一。针对DR数据集中样本类间差异小和类分布不均衡等制约分级性能提高的问题,本文提出一种融合注意力线性特征多样化(fusion of attention linear feature diversification,FALFD) 的分级算法。该算法首先用改进的Res2Net残差网络作为模型骨干来增大感受野,进一步提高网络捕捉特征信息的能力；其次引入自适应特征多样化模块(adaptive feature diversification module,AFDM) 对眼底图像可分辨的微小病理特征进行识别,获得具有高语义信息的局部特征,避免单一特征区域的限制,进而提高分级准确度；再后利用双线性注意力融合模块(bilinear attention fusion module,BAFM) 增加可判别区域特征的网络权重占比；最后采用正则化焦点损失(focal loss,FL) 进一步提升算法的分类性能。在IDRID数据集上,灵敏度和特异性分别为94.20%和97.05%,二次加权系数为87.83%；在APTOS 2019数据集上,二次加权系数和受试者工作曲线下的面积分别为88.06%和93.90%。实验结果表明,本文算法在DR分级领域中具有一定的应用价值。

Abstract:

Diabetic retinopathy (DR) is currently one of the leading blinding diseases in humans.Aiming at the problems of small differences between samples and uneven class distribution in DR datasets,which restrict the improvement of grading performance,this paper proposes a classification algorithm for the fusion of attention linear features diversification (FALFD).Firstly,the improved Res2Net residual network is used as the model backbone to increase the receptive field,and further improve the ability of the network to capture feature information.Secondly,the adaptive feature diversification module (AFDM) is introduced to identify the tiny pathological features that can be resolved in the fundus images,and local features with high semantic information are obtained,which avoids the limitation of a single feature region and improves the classification accuracy.Then,the bilinear attention fusion module (BAFM) is used to increase the proportion of network weights that can identify regional features.Finally,the regularized focal loss (FL) is used to further improve the classification performance of the algorithm.On the IDRID dataset,the sensitivity and specificity are 94.20% and 97.05%,and the quadratic weighting coefficient is 87.83%,respectively.On the APTOS 2019 dataset,the quadratic weighting coefficient and the area under the receiver operating curve are 88.06% and 93.90%,respectively.The experimental results show that the algorithm has some value in the field of DR classification.