Gated Channel Attention Network for Cataract Classification on AS-OCT Image

• Published in: Neural Information Processing Vol. 13110; pp. 357 - 368
• Main Authors: Xiao, Zunjie, Zhang, Xiaoqing, Higashita, Risa, Hu, Yan, Yuan, Jin, Chen, Wan, Liu, Jiang
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: AS-OCT; Deep learning; Gated channel attention; Nuclear cataract
• ISBN: 9783030922375; 3030922375
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-92238-2_30

Nuclear cataract (NC) is the leading cause of blindness and vision impairment globally. Accurate NC classification is significant for clinical NC diagnosis. Anterior segment optical coherence tomography (AS-OCT) is a non-contact, high-resolution, objective imaging technique, which is widely used in diagnosing ophthalmic diseases. Clinical studies have shown that there is a significant correlation between the pixel density of the lens region on AS-OCT images and NC severity levels; however, automatic NC classification on AS-OCT images has not been seriously studied. Motivated by clinical research, this paper proposes a gated channel attention network (GCA-Net) to classify NC severity levels automatically. In the GCA-Net, we design a gated channel attention block by fusing the clinical priority knowledge, in which a gated layer is designed to filter out abundant features and a Softmax layer is used to build the weakly interacting for channels. We use a clinical AS-OCT image dataset to demonstrate the effectiveness of our GCA-Net. The results showed that the proposed GCA-Net achieves 94.3% in accuracy and outperformed strong baselines and state-of-the-art attention-based networks.