Multi-modality Images Analysis: A Baseline for Glaucoma Grading via Deep Learning

Abstract

Glaucoma is one of the leading causes of blindness in humans, which is not reversible, but early detection and treatment can save helpful vision. Clinicians classify glaucoma into early, moderate, and advanced stages based on the extent of the patient’s visual field deficit. The treatment of glaucoma varies with the course of the disease. With the development of deep learning technology, more and more studies focus on the automatic diagnosis of glaucoma. Most of them are based on color fundus images or OCT images. However, there are limitations in using only one modality images to analyze glaucoma due to the complexity of glaucoma. Therefore, in this paper, two modalities of images, color fundus image and 3D OCT image provided by the GAMMA Challenge, were used to design baseline algorithms for glaucoma grading. On the preliminary dataset of the GAMMA Challenge, the kappa value of the glaucoma grading results based on the two modalities of image input were improved by 0.092 and 0.075, respectively, compared with those of the model with single fundus image and single OCT image input. And on the final datasets, the corresponding improvement were 0.029 and 0.127. At the same time, considering that optic disc changes are the main features of glaucoma, we added local information of optic disc into the input module, so that the kappa values were improved respectively by 0.075 and 0.068 in the preliminary dataset and final dataset of the model based on the images of two modalities as input. In addition, this study used an ordinal regression strategy on the classification task to increase the kappa value of the results of automatic classification of glaucoma based on multi-modality images by 0.097 and 0.050 on the preliminary and final datasets of the GAMMA Challenge.