A Simplified Deep Network Architecture on Optic Cup and Disc Segmentation

• Published in: 2020 International Joint Conference on Neural Networks (IJCNN) pp. 1 - 7
• Main Authors: Huang, Guan-Ru, Hsiang, Tien-Ruey
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.07.2020
• Subjects: Convolution; deep learning; Feature extraction; glaucoma screening; Histograms; Image segmentation; Machine learning; optic cup segmentation; optic disc segmentation; Optical imaging; Retina; Simplification
• ISSN: 2161-4407
• DOI: 10.1109/IJCNN48605.2020.9206670

Glaucoma is caused by damaged optic nerves, and can lead to permanent vision loss. The cup-to-disk ratio (CDR) is a key criterion for glaucoma diagnosis, therefore an accurate automatic segmentation of the optic disc (OD) and optic cup (OC) in retinal fundus images has become a major research topic. However, with deeper deep learning models being used to complete a fundus segmentation, the segmentation results remain acceptable only on certain datasets. In this research, a lightweight deep-learning encoder-decoder architecture, which adopts polar coordinate transformation and histogram equalization to simplify the learning complexity, is proposed for the simultaneous segmentation of OD and OC. The proposed model employs an encoder-decoder architecture consisting of a feature extraction module and a multi-output module to both simplify the model and to adjust the result through different outputs and multilabel loss functions. Experiment results demonstrate that the proposed approach outperforms other deep network models on OD and OC segmentation over the datasets REFUGE, MESSIDOR, and RIM-ONE, which contain images captured from cameras with various specifications. In addition, better results on glaucoma screening through CDR calculation were obtained on the REFUGE dataset with the proposed method.