基于Transformer和MLP的眼底血管分割算法

Retinal Blood Vessel Segmentation Based on Transformer and MLP

为了解决眼底血管分割中存在的分割效果不佳、数据过拟合和正负样本不均衡等问题,提出了一种转换器(Transformer)和多层感知机(MLP)结合的眼底血管分割算法。首先,为预防数据过拟合问题,训练图像在输入模型前会执行多种数据增强操作;其次,设计一个融合了卷积模块的Transformer组成多尺度编码器对图像进行特征提取,以此获得鲁棒的多级特征信息;最后,使用MLP结构的解码器对特征图完成像素级的分类。为解决正负样本不均衡的问题,引入了Tversky损失和二进制交叉熵损失的组合损失函数。所提算法在多个数据集上都取得了良好的实验结果,优于现有的其他网络模型算法。

To solve the problem of poor segmentation effect,data over-fitting,and imbalance of positive and negative samples in fundus blood vessel segmentation,a retinal blood vessel segmentation algorithm based on transformer architecture(Transformer)and multilayer perceptron(MLP)is proposed.First,data augmentation is used on training images to prevent over-fitting.Then,several transformers fused with convolution modules are used as a robust encoder to gain multi-scale feature information.Finally,a decoder consisting of MLP is adopted to complete pixel-level classification on a feature map.In addition,the combination of Tversky loss and binary cross-entropy loss is applied to solve the sample imbalance problem.Experiential results on various datasets indicate that the proposed algorithm has a good performance,which is better than other existing algorithms.