摘要目的 研究基于迁移学习的糖尿病视网膜病变(DR)诊断算法在小样本训练数据集中的应用.方法 采用广东省肇庆市高要区人民医院拍摄的4465幅彩色眼底照片作为完整数据集.使用固定预训练参数和微调预训练参数的模型训练策略作为迁移学习组,将其与非迁移学习的随机初始化参数的策略对比,并将这3种策略应用在ResNet50、Inception V3和NASNet 3种深度学习网络的训练上.此外,从完整数据集中随机划分出小样本数据集,研究训练数据的减少对不同训练策略的影响.采用诊断模型的准确率和训练时间分析不同训练策略的效果.结果 取不同网络架构中的最优结果.微调预训练参数策略取得的模型准确率为90.9％,高于固定预训练参数策略的88.1％及随机初始化参数策略的88.4％.固定预训练参数策略的训练所需时间为10 min,少于微调预训练参数策略的16 h及随机初始化参数策略的24 h.在训练数据减少后,随机初始化参数策略得到的模型准确率平均下降8.6％,而迁移学习组准确率平均下降2.5％.结论 结合迁移学习中的微调策略和NASNet架构的新型识别算法在小样本数据集下仍保持高准确率,具有高度的鲁棒性,可用于DR的有效筛查.

abstractsObjective To investigate a diabetic retinopathy ( DR ) detection algorithm based on transfer learning in small sample dataset. Methods Total of 4465 fundus color photographs taken by Gaoyao People ' s Hospital was used as the full dataset. The model training strategies using fixed pre-trained parameters and fine-tuning pre-trained parameters were used as the transfer learning group to compare with the non-transfer learning strategy that randomly initializes parameters. These three training strategies were applied to the training of three deep learning networks:ResNet50,Inception V3 and NASNet. In addition,a small dataset randomly extracted from the full dataset was used to study the impact of the reduction of training data on different strategies. The accuracy and training time of the diagnostic model were used to analyze the performance of different training strategies. Results The best results in different network architectures were chosen. The accuracy of the model obtained by fine-tuning pre-training parameters strategy was 90. 9％,which was higher than the strategy of fixed pre-training parameters (88. 1％) and the strategy of randomly initializing parameters ( 88. 4％) . The training time for fixed pre-training parameters was 10 minutes,less than the strategy of fine-tuning pre-training parameters ( 16 hours ) and the strategy of randomly initializing parameters (24 hours). After the training data was reduced,the accuracy of the model obtained by the strategy of randomly initializing parameters decreased by 8. 6％ on average,while the accuracy of the transfer learning group decreased by 2. 5％ on average. Conclusions The proposed automated and novel DR detection algorithm based on fine-tune and NASNet structure maintains high accuracy in small sample dataset,is found to be robust,and effective for the preliminary diagnosis of DR.