Self-speculation of clinical features based on knowledge distillation for accurate ocular disease classification. (May 2021)

• Record Type: Journal Article
• Authors: He, Junjun; Li, Cheng; Ye, Jin; Qiao, Yu; Gu, Lixu
• Is Part Of: Biomedical signal processing and control. Volume 67(2021)
• Publication Date: 2021-05
• Subjects: Ocular disease classification -- Knowledge distillation -- Clinical feature self-speculation -- Multi-label annotation; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102491
• Languages: English
• ISSNs: 1746-8094

Abstract

Graphical abstract: Highlights: A method to incorporate clinical features into ocular disease classification models. A knowledge distillation-based strategy to enhance network performance. Significantly improved imaging-based classification performance. Abstract: Ocular diseases can lead to irreversible vision impairment if not treated timely. Various imaging techniques have been developed to aid in the detection of ocular diseases, including the widely employed color fundus photography. Nevertheless, early-stage ocular diseases are difficult to be accurately diagnosed because of the few visible symptoms, and automatic ocular disease classification based only on imaging is extremely challenging. In this paper, we propose a knowledge distillation-based method to improve the performance of imaging-based automatic ocular disease classification models. Specifically, two deep neural networks are optimized sequentially. A teacher network is trained that can exploit the information from inputs of both color fundus photographs and radiologists provided clinical features. Then, through distilling the knowledge of the teacher model, a student network is learned that can self-speculate the clinical features-relevant information from the sole inputs of images. Extensive experiments validate that our student model can largely recover the performance of the teacher model and thus, the proposed method can significantly enhance the imaging-based ocular disease diagnosis without the reliance on clinical features.