Classification and segmentation of OCT images for age-related macular degeneration based on dual guidance networks

Age-related macular degeneration (AMD) is one of the main causes of visual impairment in elderly people, with drusen and choroidal neovascularization (CNV) being two characterizing types of lesions. Based on optical coherence tomography (OCT), image classification can be used in AMD diagnosis, while image segmentation is necessary for quantitative assessment of the lesion area. In this paper, we propose a deep learning framework exploiting dual guidance between the two tasks. Firstly, a complementary mask guided convolutional neural network (CM-CNN) is proposed to perform classification of OCT B-scans with drusen or CNV from normal ones, where the guiding mask is generated by the auxiliary segmentation task. Secondly, a class activation map guided UNet (CAM-UNet) is proposed to achieve segmentation of drusen and CNV lesions, using CAM output from the CM-CNN. Tested on a subset of the public UCSD dataset, and compared with five classification networks, four segmentation networks, and three multi-task networks, the proposed dual guidance network has achieved higher accuracy both in classification and segmentation. The classification accuracy reaches 96.93% and the Dice coefficient for segmentation reaches 77.51%. Results on an extra dataset for detection of macular edema and segmentation of retinal fluids further show the generalizability of the proposed model.