Zhou Chengfeng, Ye Juan, Wang Jun, Zhou Zhiyong, Wang Linyan, Jin Kai, Wen Yaofeng, Zhang Chun, Qian Dahong
School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China.
Department of Ophthalmology, the Second Affiliated Hospital of Zhejiang University, Hangzhou, China.
Biomed Opt Express. 2022 Mar 11;13(4):2018-2034. doi: 10.1364/BOE.450543. eCollection 2022 Apr 1.
Convolutional neural networks (CNNs) are commonly used in glaucoma detection. Due to the various data distribution shift, however, a well-behaved model may be plummeting in performance when deployed in a new environment. On the other hand, the most straightforward method, data collection, is costly and even unrealistic in practice. To address these challenges, we propose a new method named data augmentation-based (DA) feature alignment () to improve the out-of-distribution (OOD) generalization with a single dataset, which is based on the principle of feature alignment to learn the invariant features and eliminate the effect of data distribution shifts. creates two views of a sample by data augmentation and performs the feature alignment between that augmented views through latent feature recalibration and semantic representation alignment. Latent feature recalibration is normalizing the middle features to the same distribution by instance normalization (IN) layers. Semantic representation alignment is conducted by minimizing the Topk NT-Xent loss and the maximum mean discrepancy (MMD), which maximize the semantic agreement across augmented views from individual and population levels. Furthermore, a benchmark is established with seven glaucoma detection datasets and a new metric named mean of clean area under curve () for a comprehensive evaluation of the model performance. Experimental results of five-fold cross-validation demonstrate that can consistently and significantly improve the out-of-distribution generalization (up to +16.3% ) regardless of the training data, network architectures, and augmentation policies and outperform lots of state-of-the-art methods.
卷积神经网络（CNNs）常用于青光眼检测。然而，由于各种数据分布偏移，一个表现良好的模型在部署到新环境中时性能可能会急剧下降。另一方面，最直接的方法——数据收集，成本高昂，在实际中甚至不现实。为了应对这些挑战，我们提出了一种名为基于数据增强（DA）的特征对齐（）的新方法，以利用单个数据集提高分布外（OOD）泛化能力，该方法基于特征对齐原理来学习不变特征并消除数据分布偏移的影响。通过数据增强为样本创建两个视图，并通过潜在特征重新校准和语义表示对齐在这些增强视图之间进行特征对齐。潜在特征重新校准是通过实例归一化（IN）层将中间特征归一化为相同分布。语义表示对齐是通过最小化Topk NT-Xent损失和最大均值差异（MMD）来进行的，这在个体和总体层面上最大化了增强视图之间的语义一致性。此外，利用七个青光眼检测数据集建立了一个基准，并提出了一个名为曲线下清洁区域均值（）的新指标，用于全面评估模型性能。五折交叉验证的实验结果表明，无论训练数据、网络架构和增强策略如何，都能持续且显著地提高分布外泛化能力（高达+16.3%），并优于许多现有先进方法。
