|
Current issue
Archive
Videos
Articles in press
About the journal
Supplements
Editorial board
Reviewers
Abstracting and indexing
Subscription
Contact
Instructions for authors
Publication charge
Ethical standards and procedures
Editorial System
Submit your Manuscript
|
1/2020
vol. 122 abstract:
Original paper
A new platform designed for glaucoma screening: identifying the risk of glaucomatous optic neuropathy using fundus photography with deep learning architecture together with intraocular pressure measurements
Anna Zaleska-Żmijewska
1, 2, 3
,
Jacek P. Szaflik
1, 2, 3
,
Paweł Borowiecki
3
,
Katarzyna Pohnke
4
,
Urszula Romaniuk
4
,
Izabela Szopa
4
,
Jacek Pniewski
4
,
Jerzy Szaflik
3
KLINIKA OCZNA 2020, 122, 1: 1–6
Online publish date: 2020/04/10
View
full text
Get citation
ENW EndNote
BIB JabRef, Mendeley
RIS Papers, Reference Manager, RefWorks, Zotero
AMA
APA
Chicago
Harvard
MLA
Vancouver
Aim of the study
To develop a platform designed for glaucoma screening, based on deep learning algorithms, for the diagnosis of glaucomatous optic neuropathy from colour fundus images and intraocular pressure, not requiring medical staff. Material and methods A modular platform for glaucoma screening is developed which uses classifiers that independently evaluate the parameters of the visual system. The fundus image classifier is based on trainable mathematical models, while an intraocular pressure classifier is a threshold classifier. Performance analysis is conducted in terms of the statistical parameters: sensitivity, accuracy, precision, and specificity. Glaucoma images were classified by two experts. The cut-off of vertical cup to disc ratio (vCDR) for glaucoma was set at ≥ 0.7. In the training stage 933 healthy and 754 glaucoma images were used. If the intraocular pressure (IOP) was ≥ 24 mmHg in at least one eye the patient was classified in the glaucoma category independently of the fundus image category. During the training stage the following parameters of the image classifier were achieved: sensitivity 0.82 and specificity 0.63. Results and conclusions For the test data from two campaigns were used (total 1104 fundus images). For the image classifier, sensitivity 0.73 and specificity 0.83 were obtained for the first campaign, while sensitivity 0.84 and specificity 0.67 were obtained for the second campaign. The final achieved parameters of the platform are: sensitivity 0.79 and specificity 0.67 for the first campaign, sensitivity 0.92 and specificity 0.42 for the second campaign. The results are in accordance with other studies and the platform proved its usability and good performance. keywords:
glaucoma, artificial intelligence, image classification, screening, deep learning |
|
POLSKI
