基于混合注意力的早产儿视网膜病变分类方法

Classification of retinopathy of prematurity based on mixed attention

急性进展性后极部早产儿视网膜病变(aggressive posterior retinopathy of prematurity,AP-ROP)的患儿若不能及时得到诊断,易失去最佳治疗时间,导致患儿视网膜完全脱离,并最终失明.为更好地从ROP病例中诊断出AP-ROP,提出一种基于混合注意力网络(mixed attention network,MANet)的半监督学习方法.该方法分别通过低级数据增强器和高级数据增强器获得未标记训练数据的两种不同级别的增强数据.先利用低级数据增强器获得的增强数据生成未标记数据的伪标签,所得伪标签进一步被用作通过高级数据增强器获得的增强数据的训练标签.在设计的分类器模型中,将自注意力、空间注意力和通道注意力相结合,增强了分类器的特征提取能力.实验结果表明,本研究方法可以用很少的标记数据实现能够与全监督学习相媲美的性能,从而实现早产儿视网膜病变的自动检测,辅助医生进行ROP和AP-ROP的早期筛查.

Children with aggressive posterior retinopathy of prematurity(AP-ROP) are prone to lose the best treatment time if they are not diagnosed in time, leading to complete retinal detachment and eventual blindness. In order to diagnose this special type of AP-ROP from ROP cases more accurately, we propose a semi-supervised learning method based on mixed attention network. We obtain two different data representations of unlabeled training data through a weak enhancer and a strong enhancer, respectively. The enhanced data obtained by the weak enhancer is used to generate pseudo labels of the unlabeled data, which will be further used as the supervised training labels of enhanced data obtained by the strong enhancer. In the devised model, we consider the self-attention, spatial attention and channel attention together to enhance the capabilities of feature extraction.Experimental results show that the proposed method can achieve performance comparable to fully-supervised learning with few labeled data, thereby realizing automatic detection of retinopathy of prematurity, and assisting doctors in the early screening of ROP and AP-ROP.