Purchase this article with an account.
Fabio Scarpa, Alexa Berto, Alessia Colonna, Alberto Scarpa; Development of an automated analysis for glaucoma screening of videos acquired with smartphone ophthalmoscope. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4542.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Widespread screening is crucial for early diagnosis and treatment of glaucoma. The development of smartphone-based ophthalmoscopes represents a resource, since they are low cost and portable solutions able to imaging the optic disc easily. Acquired images have a field of view and a quality lower than conventional fundus cameras images. Therefore, standard image processing algorithms are not designed for this new type of images. Thus, we propose a completely automated analysis of videos acquired with the D-EYE smartphone-based ophthalmoscope, capable to assist ophthalmologists in performing diagnosis. This analysis is going to be integrated and tested in the See Far European Project (#826429).
1000 frames were selected from 30 videos (from different healthy subjects) acquired with the D-EYE system. The optic disc was manually segmented. These images (1080x1920 px resized to 512x1024 px) were used as training set for a u-shaped convolutional neural network (U-Net), with 4 blocks for both the encoder and decoder path to segment the optic disc. After training, the U-Net analyzes each frame of a video. Optic disc area and focus are used to select the best frame. Then, the best frame is cropped around the optic disc, resized to 512x512 px, and the cup segmentation is performed by a second U-Net. It was trained on the 750 images of the RIGA dataset, cropped around the optic disc, resized to 512x512 px, and blurred with Gaussian filters so as to have same quality of images from D-EYE system. Finally, the algorithm derives the VCDR (vertical cup-to-disc ratio).
On 5 healthy subjects and 5 subjects with glaucoma, a perfect agreement with manual analysis was obtained in the best frame selection and an accuracy ≥96% was obtained for both disc and cup segmentation. Finally, the computed VCDR denotes a substantial difference between the two groups of subjects: VCDR<0.40 for healthy subjects, VCDR>0.45 for subjects with glaucoma.
We developed a proof-of-concept automated analysis of videos acquired with the D-EYE system (Fig1). The proposed analysis can provide visual and quantitative information that assist ophthalmologists. Indeed, clinicians can look at a single image rather than an entire video and can easily verify the reliability of the automated analysis. Results encourage the further development of the proposed method and its investigation on a large dataset.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only