AI-Based Screening Tool Detects Homonymous Defects

Download PDF

Canadian researchers have harnessed deep learning to create a screening tool that is both highly accurate and effective in identifying homonymous visu­al field (VF) defects on automated perimetry.¹ This tool can alert clinicians to possible defects, which are often subtle and difficult to detect, even by skilled ophthal­mologists. This work expands the boundaries of AI in ophthalmology.

Urgency. Homonymous VF defects can indicate the presence of serious intracranial pathology, including stroke, brain tumors, demyelinating disease, and meta-stasis from cancers elsewhere in the body. When detect­ed, they indicate the need for urgent neuroimaging to examine the chiasmal and retrochiasmal visual pathways for the presence of lesions.

However, timely detection of these defects may not occur, leading to significant patient morbidity and mor­tality. “Many ophthalmologists have grappled with cases where, on follow-up visits, they realize that they’d missed a subtle homonymous visual field defect,” said coauthor Edward Margolin, MD, FRCSC, Dipl. ABO, at the University of Toronto.

Study design. For this retrospective proof-of-con­cept study, the researchers considered medical records of 416 patients, 18 years or older, with known hom­onymous defects seen during a five-year period at a university-affiliated high-volume neuro-ophthalmology practice.  

In addition, they randomly collected VF tests from 820 patients seen during the same time frame. This control dataset included normal VFs and those with non-neurological VF defects, such as glaucomatous defects.

VF testing was done in both eyes by Humphrey Field Analyzer 24-2 SITA-Fast, the most commonly performed automated perimetry in ophthalmology practices. All data underwent 7-fold cross validation for training and evaluation of the proposed AI model, which the researchers dubbed Deep Homonymous Classifier (DHC).

Better than the benchmark. The DHC model proved superior to the benchmark neurological hemifield test (NHT), a score-based approach built on a mathematical rather than AI algorithm. “Our model demonstrated very high sensitivity and specificity,” Dr. Margolin commented.

Specifically, the DHC model achieved an average accuracy of 87% for detecting homonymous VF defects in previously unseen VFs. It also achieved an average recall or sensitivity of 92% and an average specificity of 88%. In contrast, the NHT achieved an average accura­cy of 66%, a recall/sensitivity of 92%, and a specificity of 47%.

The researchers noted that their DHC model is cur­rently limited by the size of the dataset. “As with all deep learning models, the proposed model will im­prove in generalizability with a larger training set and will be able to further validate the presented recall and accuracy with a larger testing set,” they wrote.

—Miriam Karmel

___________________________

1 Tan AH et al. Br J Ophthalmol. Published online Aug. 3, 2022.

___________________________

Relevant financial disclosures: Dr. Margolin—None.

For full disclosures and the disclosure key, see below.

• Impact of Sleep Apnea on Retinal Microvasculature

• Guidelines for Perioperative Management of Antithrombotics

• Assessing Visual Issues Following Concussion