一种多任务学习结合U-Net的微动脉瘤图像分割方法

A multi-task learning combined with U-Net’s microaneurysm image segmentation method

微动脉瘤是检测糖尿病初期视网膜病变的关键病灶,针对微动脉瘤图像分割问题,提出了多种图像预处理技术与多任务学习网络相结合的方法。该方法使用了多种图像预处理技术,在UNet中引入了注意力机制,并将微动脉瘤图像分割作为主任务,微动脉瘤存在性检测作为副任务,利用多任务学习结合U-Net来提升主任务分割效果。在国际公开数据集上进行实验,获得了AUC为9.48%以及AP为51.40%的结果,与单任务图像分割相比,AP值提升了3.82%,由实验结果可知该方法能够提升微动脉瘤的分割效果。

Microaneurysm is the key lesions in the detection of early Diabetic Retinopathy,aiming at the problem of microaneurysm detection in fundus images,a method combining multi-task learning network and various image preprocessing is proposed.This method uses a variety of image preprocessing techniques,introduces an attention mechanism into U-Net,and uses microaneurysm image segmentation as the main task,and microaneurysm presence detection as the secondary task,it uses multi-task learning combined with U-Net to improve the main task segmentation effect.Experiments were conducted on an internationally available dataset,also achived results with AUC of 99.48%and AP of 51.40%,compared with single-task image segmentation,the AP value increased by 3.82%,the experimental results show that this method can improve the segmentation effect of microaneurysms.