摘要
目的:探索不均衡小样本的眼底彩照病理性近视自动检测方法,使得深度学习模型能在病例量较小且分布不均衡的情况下取得同样有效的结果,推动深度学习方法在中小规模眼科临床中的应用。方法:选取南京医科大学附属逸夫医院眼科中心门诊的466只近视眼底彩照作为小样本病例数据,以目前最新的TNT神经网络模型(Transformer-iN-Transformer)作为基准,采用对不均衡小样本病例数据进行数据增强的方法,建立小样本的眼底彩照病理性近视自动检测模型,并与传统的大规模训练模型性能进行对比。结果:合理的数据增强方法能为基准模型带来显著的性能提升,最终建立的眼底彩照病理性近视自动检测模型在不均衡小样本上的准确率为(98.09±0.89)%,敏感性为(93.33±6.09)%,特异性为(88.14±7.92)%,AUC值为0.99,达到了传统大规模训练模型的性能。结论:在病例量较小且分布不均衡的情况下,合理的数据增强方法能显著改善深度学习模型的性能,对病理性近视检测具有与大规模训练模型同等的准确率、敏感性和特异性,这些结果将有助于进一步推广对病理性近视的自动筛查,并在随后保护患者免受由病理性近视眼底病变引起的低视力和失明。
Objective:To explore an automatic detection method of pathologic myopia in color fundus images with imbalanced small samples,so that the deep learning model can achieve competitive results and promote the application of deep learning methods in small and medium-scale ophthalmology clinics.Methods:The color fundus images of 466 myopic eyes from the outpatient department of the Sir Run Run Hospital Nanjing Medical University were selected as the small sample dataset,and the state-of-the-art neural network model TNT(Transformer-iN-Transformer)was used as the baseline.Data augmentation methods were used to build an imbalanced small sample pathologic myopia automatic detection model,and to compare the performance with the traditional large-scale models.Results:Proper data augmentation methods can bring significant performance enhancement to the benchmark model.The average accuracy of the detection model was(98.09±0.89)%,the sensitivity was(93.33±6.09)%,the specificity was(88.14±7.92)%,the AUC value was 0.99,which was a competitive performance of traditional large-scale models.Conclusion:In the case of lacking of data and imbalanced distribution,proper data augmentation methods can significantly improve the performance of deep learning models,and had the competitive accuracy,sensitivity and specificity as large-scale models for pathologic myopia detection.These results will help the further promotion of automatic pathologic myopia detection and subsequently protect patients from low vision and blindness caused by pathological myopic fundus lesions.
作者
刘懿
孟凡杰
章超伟
刁鹏飞
LIU Yi;MENG Fanjie;ZHANG Chaowei(Department of Ophthalmology,Sir Run Run Hospital Nanjing Medical University,Nanjing City,Jiangsu Province 211100;不详)
出处
《医学理论与实践》
2023年第15期2538-2541,共4页
The Journal of Medical Theory and Practice
基金
南京医科大学附属逸夫医院重点学科基金(YFZDXK02-4)。
关键词
不均衡小样本
病理性近视
自动检测
数据增强
Imbalanced small samples
Pathologic myopia
Automatic detection
Data augmentation
