DSLN: Dual-tutor student learning network for multiracial glaucoma detection

Accurate early glaucoma detection is crucial to prevent further vision loss. However, using the off-the-shelf models against fundus image datasets of different races may lead to degraded performance due to domain shift. To address the issue, this paper proposes a dual-tutor student learning network (DSLN) for multiracial glaucoma detection. The proposed DSLN consists of an inter-image tutor, an intra-image tutor, student model and backbone network, which combines the advantages of domain adaptation and semi-supervised learning. The inter-image tutor uses CycleGAN for style transfer to reduce the appearance differences between labeled source domain and labeled target domain images, and transfers the learned knowledge to the student model by minimizing knowledge distillation loss. The intra-image tutor adopts the exponential moving average to leverage the unlabeled target domain and transfers the knowledge to the student model by minimizing prediction consistency loss. Moreover, the student model not only directly learns knowledge from the labeled target domain images, but also learns the intra-image knowledge and inter-image knowledge transfer by two tutors. Furthermore, the backbone integrates the context features of the local optic disc region and global fundus image via modified ResNet50. We conduct extensive experiments on three scenarios constructed from nine public fundus image datasets of three races. Comprehensive experimental results show that the proposed DSLN framework outperforms the state-of-the-art models and has good robustness and generalization: it can effectively overcome domain shift and accurately detect glaucoma from multi-ethnic fundus images.