Abstract: In recent years, convolutional neural networks (CNNs) have been widely used for automatic age-related cataract classification. However, incorporating clinical prior knowledge of age-related cataracts into CNN design improves the classification performance and the interpretability of the decision-making process of CNNs, which has been less studied. To this problem, this paper proposes a clinical-feature recalibration attention network (CFANet) to classify age-related cataract severity levels automatically. In the CFANet, a simple yet effective clinical feature recalibration attention (CFA) block is designed to fuse clinical features adaptively by setting relative weights, aiming to highlight significant channels and suppress redundant ones. This paper conducts extensive experiments on a clinical AS-OCT image dataset of nuclear cataract and a public eye image dataset to verify the effectiveness of CFANet. The results show that CFANet outperforms advanced baselines by above 3.54 percentage points of accuracy on the clinical AS-OCT image dataset, such as squeeze-and-excitation network (SENet), efficient channel network (ECANet), style-based recalibration module (SRM). And the results on the public eye dataset also show that compared with strong attention-based CNNs and published works, proposed method obtains over 1 percentage point improvement. Moreover, this paper also uses visualization methods to analyze clinical feature weights and channel attention weights to enhance the interpretability of the decision-making process for proposed method.

Key words: age-related cataract classification, anterior segment optical coherence tomography (AS-OCT), clinical-feature recalibration attention, interpretability, convolutional neural network (CNN), visualization

摘要： 近年来，卷积神经网络（convolutional neural network，CNN）模型已经被广泛用于年龄相关性白内障自动分类任务，然而，鲜有研究工作将年龄相关性白内障的临床先验知识注入卷积神经网络架构设计中，以此来提高年龄相关性白内障的分类效果和改善模型决策过程的可解释性。提出了一种临床特征校准注意力网络（clinical feature recalibration attention network，CFANet）模型用于自动识别年龄相关性白内障严重级别。在CFANet中，设计了一个简单且有效的临床特征校准注意力模块（clinical feature recalibration attention block，CFA），其不仅能对不同临床特征类型进行自适应地加权融合，还通过门控操作符来突出重要通道和抑制不重要通道。在一个核性白内障的眼前节光学相干断层成像影像（anterior segment optical coherence tomography，AS-OCT）数据集和一个公开眼科影像数据集上进行了充分实验，实验结果表明，相较于squeeze-and-excitation network（SENet）、efficient channel network(ECANet)、style-based recalibration module（SRM），CFANet在AS-OCT数据集上的分类准确率至少提升了3.54个百分点，同时在公开的眼科影像数据集上的分类结果比先进的神经网络模型和已发表的研究工作提升了1个百分点以上。此外，还通过可视化方法分析临床特征的权重分布和通道的注意力权重分布来提高该文模型决策过程的可解释性。

关键词: 年龄相关性白内障分类, 眼前节光学相干断层成像, 临床特征校准注意力模块, 可解释性, 卷积神经网络, 可视化