Abstract
Purpose :
To develop a deep learning framework that can automatically identify clinically relevant features on glaucoma fundus images.
Methods :
Fundus photographs from 267 normal eyes and 160 eyes with glaucoma were included. We developed a deep learning model based on the pre-trained NASNet and used heat map technique to assess parts of the fundus image that were driving the classification, thus allowing localization of clinically relevant objects on retinal fundus images. After training the model, all 427 fundus images were used as input to the proposed model (based on a deep pre-trained classifier) consisting of the region of interest on fundus photographs. The clinical diagnosis labels of fundus images were validated by a glaucoma specialist and the outcome of deep learning was assessed by experts to assure clinical relevance.
Results :
The accuracy of the method in discriminating normal eyes from eyes with glaucoma was 92%. The validation accuracy on an independent dataset of 455 images was 90%. Among fundus images that had been classified to glaucoma group, we observed that deep learning had identified significant features mostly in the superior/inferior peripapillary regions, within the optic nerve head, as well as in their pattern of large blood vessel structure.
Conclusions :
We developed a deep learning model based on pre-trained parameters that was able to detect clinically relevant glaucoma features from fundus images with high accuracy. This approach could be useful in glaucoma clinics as well as in general practice settings as an assistive tool for screening glaucoma in the absence of glaucoma clinicians. Validation of our findings in an independent cohort with larger number of fundus images is required.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.