融合密集空洞注意力金字塔和多尺度的视网膜病变分割

Diabetic retinopathy segmentation using dense dilated attention pyramid and multi-scale features

针对糖尿病视网膜病变(DR)分割任务中病变区域多尺度特征难以学习、边界模糊等问题,提出一种改进的U型多病变分割模型DDAPNet。首先,对DR图像进行Patch处理,使模型更好地捕捉病变的局部特征;其次在主干特征提取后引入重新设计的密集空洞注意力金字塔(DDAP)模块,扩大感受野,解决病变边界模糊问题;同时采用金字塔切分注意力进行特征增强,然后将二者进行特征融合;最后在跳跃连接中嵌入改进的残差注意力模块,降低浅层冗余信息的干扰。在数据集和医院真实数据集上进行联合验证,实验结果表明,相较于基础模型,DDAPNet模型对微动脉瘤、出血点、软渗出DDR物和硬渗出物的分割在Dice系数上分别提高了4.31%、2.52%、3.39%、4.29%,在mIoU上分别提高了1.80%、2.24%、4.28%、1.98%。该模型对病灶边缘的分割更为连续和平滑,有效提升了软渗出物等视网膜病变的分割性能。

An improved U-shaped multi-lesion segmentation model,namely dense dilated attention pyramid UNet(DDAPNet),is proposed to overcome the difficulty in learning multi-scale features and address the issue of blurry boundaries in diabetic retinopathy(DR)segmentation task.DR images are treated with Patch processing to enhance the model's ability to capture local lesion features.After backbone feature extraction,a redesigned dense dilated attention pyramid module is introduced to expand the receptive field and address the issue of blurry lesion boundaries;and simultaneously,pyramid split attention module is used for feature enhancement;and then,the features output by the two modules are fused.Additionally,an improved residual attention module is embedded within skip connections to reduce interference from shallow redundant information.The joint validation on DDR dataset and real dataset from a specific hospital shows that compared with the original model,DDAPNet model improves the Dice similarity coefficient for segmentations of microaneurysms,hemorrhages,soft exudates and hard exudates by 4.31%,2.52%,3.39%and 4.29%,respectively,and increases mean intersection over union by 1.80%,2.24%,4.28%and 1.98%,respectively.The proposed model makes the segmentation of lesion edges smoother and more continuous,notably enhancing the segmentation performance for conditions like soft exudates in retinal lesions.