Analyse The Clinical Validation Of Artificial Intelligence (Ai) For Detecting Glaucoma.

Detection of glaucoma in time is very important for sight-saving glaucoma treatment and surgery.

To investigate the potential of developing artificial intelligence (AI) for glaucoma detection in an anonymised fundus photographic dataset obtained in a clinical environment. 

Retrospective, analytical descriptive. 

The Deepdee AI algorithm model (AI model) was trained based on the AIROGS (Artificial Intelligence for Robust Glaucoma Screening Challenge) training set using Fastai/PyTorch. This AI model was then tested on two different datasets: 1. Clinical dataset and 2. AIROGS test set. The clinical dataset consists of 217 posterior pole photographs (from 171 patients), 106 of which were taken on eyes with OAG and 111 on eyes without OAG (control group). Eyes with OAG had visual field defects on standard HFA 30-2 w/w testing; they were stratified by severity using simplified Hodapp classification¹ into early, moderate and advanced glaucoma. All patients had undergone a full ophthalmological examination and have as well pseudophakia, DR, and/or AMD.  

The results of the validation of the AI were distributed as follows:

The AIROGS test set: accuracy - 0.97, sensitivity - 0.98, specificity - 0.97, PPV - 0.97, NPV - 0.98.

The Clinical dataset: accuracy - 0.82, sensitivity - 0.72, specificity - 0.91, PPV - 0.88, NPV - 0.78.

PPV - positive predictive value

NPV - negative predictive value

The distribution of the results in the OAG group in accordance with field loss was also analysed.

FN appeared mostly in the group of early glaucomatous VF loss MD < 6 dB, followed by moderate and advanced glaucomatous loss. 

FN appeared in the eyes of patients with the other pathology, disc abnormalities such as neovascularization, hemorrhages, or proliferation in the optic disc area.

Validation of the AI model on two datasets - the AIROGS test set and the Clinical dataset, showed the following: accuracy, sensitivity, PPV, NPV are higher in the AIROGS test set than in the Clinical dataset.

The AIROGS dataset was created to develop a robust AI for screening the populations at large under real-life circumstances. Fundus images with non-glaucomatous disc pathology can be expected to lead to poorer AI results.

The AI model needs real-world validation and requires further investigation on a larger group of patients in screening population models for more statistically significant analysis. 

AI analysis of fundus retinal images is a promising method for glaucoma screening and promises to reduce manual testing.