基于PL-EfficientNet的糖尿病视网膜病变检测研究

Diabetic retinopathy detection research based on PL-Efficient Net

糖尿病视网膜病变(DR)早期检测可使患者得到及时诊治,具有重要意义。针对现有检测算法中无标签眼底图像数据利用率不足、DR不同病变等级之间特征差异小、模型难以捕捉分类微细节而造成诊断效率欠佳的问题,提出一种基于伪标签技术及高效网络PL-Efficient Net检测算法,该算法利用迁移学习加载预训练Efficient Net系列模型,微调后采用伪标签技术引入无标签数据,结合标签数据二次训练得到最优结果。通过在Kaggle竞赛Diabetic Retinopathy Detection、Aptos 2019Blindness Detection公开数据集上进行实验,验证所提算法平方加权(Kaggle)值达到了0.918,相比其他先进算法,眼底图像数据利用率以及诊断效率均得到一定提升。该算法在临床上对眼科医生也可起到积极的辅助作用。

Early detection of Diabetic Retinopathy(DR) is of great significance for patients to receive timely diagnosis and treatment.In view of the insufficient utilization of the unlabeled fundus image data of the existing detection algorithms,the small feature difference between different lesion levels in DR,and the difficulty of the model to capture the classification of micro-details resulting in poor diagnosis efficiency that a new method based on Pseudo-Label technology and high-efficiency network PL-Efficient Net detection algorithm is proposed. The algorithm uses transfer learning to load the pre-trained Efficient Net series models,after fine-tuning,uses Pseudo-Labels to introduce unlabeled data,combined with labeled data for secondary training to obtain the best results. Through experimental verification on the public datasets of the Kaggle competition Diabetic Retinopathy Detection and Aptos 2019 Blindness Detection,the proposed algorithm has a square weighted(Kaggle) value of 0.918.Compared with other advanced algorithms, the fundus image data utilization rate and diagnosis efficiency have been improved. This algorithm can also play an active role in assisting ophthalmologists in clinical practice.