AI and glaucoma: a project exploring image analysis

Glaucoma can damage the optic nerve and cause vision loss. The WHO’s global vision-impairment figure is not specific to glaucoma: at least 2.2 billion people have near or distance vision impairment, and in at least 1 billion cases it could have been prevented or remains unaddressed.

WHO: vision impairment

Editorial correction, 25 October 2026: the units in the original figure and its attribution to glaucoma have been corrected. The project retains its original 2022 publication context.

This article describes a research proposal using eye images and neural networks to explore support for clinical analysis. Results on one dataset do not establish a diagnosis or demonstrate that a tool works in other settings or populations.

How would image and software processing be carried out?

Once we acquire the different fundus images, these will be used for training the development of the eye. neural network. All of these, taking into account that they have previously been duly analyzed by specialists and obtained from various medical centers, are classified as:

-Suspected glaucoma disease
-Healthy patient
-Total sum of all these

(Previously discarding on the other hand all those that had been defective)

Jupyter Notebook

In the case of the study carried out at the Salesian Polytechnic University, it was used Python with the interpreter Jupyter Notebook that being an extension of Python It did not require additional downloads of additional programs on the PC or any type of license. Therefore, once this procedure was achieved, the images provided by various medical centers that made up the database were separated into different folders, these being separated again into two others: one with the classification of the healthy patients and others with suspected glaucoma.

The images, extracted from an optical coherence tomograph, were duly cropped until the optic disc was isolated since this is the region of interest for the diagnosis. To do this, an algorithm was created with two very clear objectives: first, that the image was normalized to dimensions similar to those of a square and second, that these also had centered the location of region of interest.

Algorithm used to crop images.

Subsequently, in order to process these images correctly, they were all the same size to speed up the process. Next, they were classified into two different folders (suspect and healthy) to correctly train the convolutional network and reserving those that would be used for the confusion matrix test.

Search for optimal solutions

The proposal shows how to prepare images and train a classifier. Assessing its clinical usefulness would require independent evaluation, representative data and review by healthcare professionals. Performance depends on data quality, model design and how training and test sets are separated; high accuracy alone does not establish clinical safety.

If you want to be part of the search for optimal solutions, do not hesitate to start or continue your training in Artificial Intelligence with us. Check out the Saturdays.AI courses and choose the one you like the most in our website.

Bibliography

1.-Buri Abad, Edison Arturo. Processing digital images of the fundus with the use of artificial intelligence to provide a support tool for the presumptive diagnosis of human glaucoma. Feb-2022.

2.-Mayo Clinic. Glaucoma. Consult at: https://www.mayoclinic.org/es-es/diseases-conditions/glaucoma/symptoms-causes/syc-20372839#:~:text=El%20glaucoma%20es%20un%20grupo,personas%20mayores%20de%2060%20a%C3%B1os.

Written by

Eva Arnall

Saturdays.AI

Saturdays.AI is an impact-focused organization on a mission to empower diverse individuals to learn Artificial Intelligence in a collaborative and project-based way, beyond the conventional path of traditional education