CAFR-CNN: coarse-to-fine adaptive faster R-CNN for cross-domain joint optic disc and cup segmentation

Glaucoma is a leading cause of blindness. Accurate and efficient segmentation of the optic disc and cup from fundus images is important for glaucoma screening. However, using off-the-shelf networks against new datasets may lead to degraded performances due to domain shift. To address this issue, in this paper, we propose a coarse-to-fine adaptive Faster R-CNN framework for cross-domain joint optic disc and cup segmentation. The proposed CAFR-CNN consists of the Faster R-CNN detector, a spatial attention-based region alignment module, a pyramid ROI alignment module and a prototype-based semantic alignment module. The Faster R-CNN detector extracts features from fundus images using a VGG16 network as a backbone. The spatial attention-based region alignment module extracts the region of interest through a spatial mechanism and aligns the feature distribution from different domains via multilayer adversarial learning to achieve a coarse-grained adaptation. The pyramid ROI alignment module learns multilevel contextual features to prevent misclassifications due to the similar appearances of the optic disc and cup. The prototype-based semantic alignment module minimizes the distance of global prototypes with the same category between the target domain and source domain to achieve a fine-grained adaptation. We evaluated the proposed CAFR-CNN framework under different scenarios constructed from four public retinal fundus image datasets (REFUGE2, DRISHTI-GS, DRIONS-DB and RIM-ONE-r3). The experimental results show that the proposed method outperforms the current state-of-the-art methods and has good accuracy and robustness: it not only avoids the adverse effects of low contrast and noise interference but also preserves the shape priors and generates more accurate contours.