基于改进U-Net的糖尿病视网膜渗出物分割

Segmentation of diabetic retinal exudates based on improved U-Net

为解决当前深度分割模型对糖尿病视网膜病变渗出物区域精准分割的问题,提出一种基于改进U-Net的糖尿病视网膜渗出物分割方法。重构编码端的特征提取器,以ResNet50预训练模型作为特征提取器,提升模型对低维特征的学习能力;结合注意力机制优化解码端上采样过程,缓解卷积特征丢失的问题,提高模型对病灶特征通道权重的学习能力;在模型解码端构造多特征尺度的融合,丰富不同尺度特征的语义信息。利用IDRID数据集和DIARETDB1数据集进行训练,训练采用Focal loss焦点损失函数,模型的特异性、灵敏度、精准率和F1-Score分别达到0.973、0.982、0.971和0.875,Dice系数达到了0.984,与其它先进算法相比,该算法分割性能较优。

To solve the problem of accurate segmentation of diabetic retinopathy exudate region using current deep segmentation model,a segmentation method of diabetic retinal exudate based on improved U-Net was proposed.The feature extractor on the encoding side was reconstructed,and the ResNet50 pre-training model was used as the feature extractor to improve the learning ability of the model for low-dimensional features.The upsampling process on the decoding side was optimized by combining the attention mechanism,which alleviated the problem of loss of convolutional features.It also effectively improved the model’s ability to learn the weights of lesion feature channels.The fusion of multi-feature scales was constructed at the decoding end of the model to enrich the semantic information of features at different scales.Using the IDRID dataset and DIARETDB1 dataset for training,the Focal loss function was used in the training,the specificity,sensitivity,accuracy and F1-Score of the model reach 0.973,0.982,0.971 and 0.875 respectively,and the Dice coefficient reaches 0.984.Compared with other advanced algorithms,this algorithm has better segmentation performance.