Abstract

Deep learning (DL) has recently gained increasing attention in biomedical data analytics, demonstrating robust performance and promising results. A deep network requires massive amount of data to learn meaningful patterns useful in solving complex problems. Data scarcity in medical field is a bottleneck for applying deep learning in this area. This has led to the popularity of pre-trained models, trained on huge source data to achieve reasonable accuracy in medical diagnosis even with less data in target domain. A wise choice of models trained with data similar to target data would ensure that relevant features are captured. In this work, the significance of choosing appropriate pre-trained models is demonstrated. The EyeNet model, originally trained for diagnosis of diabetic retinopathy (DR) using fundus image dataset, is used as a pre-trained model for building a convolutional neural network (CNN) – based DL architecture for glaucoma diagnosis using images from the same modality. The results are compared with glaucoma diagnosis using different pre-trained models that are less relevant to the problem considered. Different experiments including fine-tuning and transfer learning were performed. Results were validated using the benchmark Rim-one dataset. The EyeNet model outperformed all other models, achieving a maximum accuracy of 89% with transfer learning using support vector machines (SVM) combined with principal component analysis (PCA) for dimensionality reduction.

中文翻译：

---

EyeNet 模型在青光眼诊断中的性能评估

摘要

深度学习 (DL) 最近在生物医学数据分析中受到越来越多的关注，展示了强大的性能和有希望的结果。深度网络需要大量数据来学习对解决复杂问题有用的有意义的模式。医学领域的数据稀缺是深度学习在该领域应用的瓶颈。这导致了预训练模型的流行，这些模型在大量源数据上进行训练，即使目标域中的数据较少，也能在医疗诊断中达到合理的准确性。明智地选择使用与目标数据相似的数据训练的模型将确保捕获相关特征。在这项工作中，证明了选择合适的预训练模型的重要性。EyeNet 模型，最初训练用于使用眼底图像数据集诊断糖尿病视网膜病变 (DR)，用作构建基于卷积神经网络 (CNN) 的预训练模型，用于使用来自相同模态的图像进行青光眼诊断的深度学习架构。使用与所考虑问题不太相关的不同预训练模型将结果与青光眼诊断进行比较。进行了不同的实验，包括微调和迁移学习。使用基准 Rim-one 数据集验证了结果。EyeNet 模型优于所有其他模型，使用支持向量机 (SVM) 结合主成分分析 (PCA) 进行降维的迁移学习实现了 89% 的最高准确率。使用与所考虑问题不太相关的不同预训练模型将结果与青光眼诊断进行比较。进行了不同的实验，包括微调和迁移学习。使用基准 Rim-one 数据集验证了结果。EyeNet 模型优于所有其他模型，使用支持向量机 (SVM) 结合主成分分析 (PCA) 进行降维的迁移学习实现了 89% 的最高准确率。使用与所考虑问题不太相关的不同预训练模型将结果与青光眼诊断进行比较。进行了不同的实验，包括微调和迁移学习。使用基准 Rim-one 数据集验证了结果。EyeNet 模型优于所有其他模型，使用支持向量机 (SVM) 结合主成分分析 (PCA) 进行降维的迁移学习实现了 89% 的最高准确率。