多病种眼底疾病诊断的深度学习策略讨论

Discussion on Deep-Learning Strategies for Diagnosis of Multiple Diseases in Fundus Diseases

深度学习算法可实现精准的眼底疾病诊断,对眼底疾病的早期诊断和及时干预具有重要意义。近年来涌现出多种深度学习诊断算法,在数据扩增、图像增强、训练策略、特征提取等方面提出了改进设计,但对眼底疾病这样的多病种诊断问题,采用哪种训练策略缺乏系统的分析。从数据处理、训练测量、网络模型、注意力机制等方面分析不同策略的性能,为多种疾病诊断提供深度学习算法设计依据:独立训练后综合投票法的结果优于集中训练法,但训练成本更高;在训练集上,集中训练法的表现差异不大,但在验证集上,Res NeSt50的表现最佳,具有良好的泛化能力;单病种独立训练的二分类预测结果明显优于多病种的综合分类,表明优化分类网络结构有助于提高类别特征的分类效果。未来可考虑设计融合临床知识的专用网络结构进行特定细节增强,强化类间特征差异,使模型具有一定的可解释性,并提高诊断的准确率。这种多层融合网络设计将是眼科疾病诊断的发展方向之一。

Deep-learning algorithms can achieve the precise diagnosis of fundus diseases,which is crucial for the early diagnosis and timely intervention of these diseases.In recent years,various deep-learning diagnostic algorithms have been proposed to improve data augmentation,image enhancement,training strategies,and feature extraction.However,the systematic analysis of training strategies for the diagnosis of multiple fundus diseases is insufficient.Therefore,this study analyzes the performances of different strategies based on aspects such as data processing,training measurement,network model,and attention mechanism,thus providing a basis for the design of deep-learning algorithms to facilitate the diagnosis of various diseases.The results of independent training followed by comprehensive voting are superior to those of centralized training,although a higher training cost is incurred.On the training set,the performance differences among algorithms using centralized training are insignificant.However,on the validation set,ResNeSt50 performs the best and demonstrates its excellent generalization ability.The binary prediction results of the algorithms using single-disease independent training are significantly better than the comprehensive classification of multiple diseases,thus indicating that optimizing the classification network structure improves the classification results effect of category features.In the future,a specialized network structure that integrates clinical knowledge for specific detail enhancement shall be designed to strengthen the differences between interclass features,thus rendering the model more interpretable and improving the diagnostic accuracy.Such a multi-layered integrated network design might be one of the future directions for the diagnosis of ophthalmic diseases.