B-Scan Attentive CNN for the Classification of Retinal Optical Coherence Tomography Volumes
B-Scan Attentive CNN for the Classification of Retinal Optical Coherence Tomography Volumes
复制标题
DOI：
10.1109/lsp.2020.3000933
复制
发表时间：
2020-01-01
影响因子：
3.9
通讯作者：
Bora, Prabin Kumar
中科院分区：
文献类型：
--
作者：
Das, Vineeta;Prabhakararao, Eedara;Bora, Prabin Kumar
光学相干断层扫描（OCT）能够对视网膜组织进行3D横截面成像，并已成为眼科疾病诊断的重要工具。在临床上，眼科医生检查3D OCT体积的每个横截面图像（B扫描）以诊断视网膜病变。然而，这个过程是耗时和繁琐的。文献中的自动化方法对单个B扫描进行分类，并使用基于阈值的手动规则汇总OCT体积的诊断决策。然而，这些方法缺乏普遍性，未能纳入眼科医生的临床诊断方面。因此，在这封信中，我们提出了一种B扫描注意卷积神经网络（BACNN），它通过在分类过程中关注临床信息B扫描来模仿眼科医生的诊断方式。具体地，基于CNN的特征提取模块用于从B扫描提取空间特征表示。然后，自我关注模块根据这些特征的临床相关性聚集这些特征，以获得用于可靠诊断的有区别的高级特征向量。两个临床OCT数据集上的实验结果表明，该方法提供了更好的分类性能。该方法还证明了B扫描中的病理症状与注意力权重的相关性。
Optical coherence tomography (OCT) enables 3D cross-sectional imaging of the retinal tissues and has become an essential tool for the diagnosis of eye diseases. Clinically, the ophthalmologists examine each cross-sectional image (B- scan) of the 3D OCT volume to diagnose the retinal pathologies. However, this process is time-consuming and tedious. Automated methods in literature classify the individual B-scans and aggregate the diagnosis decision for the OCT volume using manual threshold-based rules. However, these methods lack generalizability and fail to incorporate the ophthalmologists' aspects of clinical diagnosis. Therefore, in this letter, we propose a B-scan attentive convolutional neural network (BACNN) that mimics the ophthalmologists' way of diagnosis by focusing on the clinically informative B- scans during the classification process. Specifically, a CNN based feature extraction module is used to extract spatial feature representations from the B-scans. Then, a self-attention module aggregates these features based on their clinical relevance to obtain a discriminative high-level feature vector for a reliable diagnosis. The experimental results on two clinical OCT datasets indicate that the proposed method provides improved classification performance. The method also demonstrates the correlation of the pathological symptoms in the B- scans with the attention weights.