双线性内卷神经网络用于眼底疾病图像分类

Bilinear involution neural network for image classification of fundus diseases

由于眼底图像具有复杂程度高、个体差异弱、类间距离短等特点,纯卷积神经网络(CNN)和基于注意力的网络并不能在眼底疾病图像分类任务上达到令人满意的精度。因此,采用involution算子实现了注意力双线性内卷神经网络(ABINN)模型用于眼底疾病图像分类。ABINN模型的参数量仅是传统双线性卷积神经网络(BCNN)模型的11%,并提取了眼底图像的底层语义信息和空间结构信息进行二阶特征融合,是CNN和注意力方法的有效并联。此外,提出了两种基于involution算子实现注意力计算的实例化方法:基于图块的注意力子网络(AST)和基于像素的注意力子网络(ASX),这两种方法可以在CNN的基础结构内完成注意力的计算,从而使双线性子网络能在同一个架构下训练并进行特征融合。在公开眼底图像数据集OIA-ODIR上进行实验,结果显示ABINN模型的精度为85%,比通用BCNN模型提高了15.8个百分点,比TransEye模型提高了0.9个百分点。

Due to the high complexity, weak individual differences, and short inter-class distances of fundus image features, pure Convolutional Neural Networks(CNNs) and attention based networks cannot achieve satisfactory accuracy in fundus disease image classification tasks. To this end, Attention Bilinear Involution Neural Network(ABINN) model was implemented for fundus disease image classification by using the involution operator. The parameter amount of ABINN model was only 11% of that of the traditional Bilinear Convolutional Neural Network(BCNN) model. In ABINN model, the underlying semantic information and spatial structure information of the fundus image were extracted and the second-order features of them were fused. It is an effective parallel connection between CNN and attention method. In addition, two instantiation methods for attention calculation based on involution operator, Attention Subnetwork based on PaTch(AST) and Attention Subnetwork based on PiXel(ASX), were proposed. These two methods were able to calculate attention within the CNN basic structure, thereby enabling bilinear sub-networks to be trained and fused in the same architecture.Experimental results on public fundus image dataset OIA-ODIR show that ABINN model has the accuracy of 85%, which is 15. 8 percentage points higher than that of the common BCNN model and 0. 9 percentage points higher than that of TransEye(Transformer Eye) model.