Deep feature extraction via adaptive collaborative learning for drusen segmentation from fundus images

Drusen are an early sign of non-neovascular age-related macular degeneration which is a major factor of irreversible blindness. Drusen segmentation plays a vital role in proper diagnosis and prevention of further complications. However, most of the existing drusen segmentation approaches rely on handcrafted features which are not always guaranteed to be discriminative and therefore lead to limited performance. In this paper, we propose a deep feature extraction framework for drusen segmentation. It is formulated as a deep model which can automatically extract discriminative features. Specifically, the framework is mainly composed of three components, including feature learning, loss function and classification. The effectiveness of our method lies in the fact that the deep feature learning procedures are driven by an adaptive collaborative similarity learning technique in loss function. We evaluate the framework on STARE and DRIVE datasets, and the quantitative comparison with the state-of-the-art methods in terms of sensitivity, specificity and accuracy demonstrates the superiority of the proposed method.