Artificial intelligence systems are becoming increasingly available as diagnostic aids for optometric practice. These perform well but are often ‘black-box’ systems offering little or no insight into how a decision was reached. While there is potential for artificial intelligence to improve patient outcomes, clinicians without training in computer science may find it difficult to ascertain whether these technologies are suitable for their practice, or how they should be used. This review provides an overview of how artificial intelligence systems work in optometry, their strengths, weaknesses, and regulatory considerations. A checklist is provided for appraising a system, covering regulatory approvals, ascertaining what the system can and cannot do, how it can be used in practice, whether it is suitable for the clinical population, and whether the outputs can be explained. Artificial intelligence has the potential to improve accuracy and efficiency in many areas of optometry if used correctly, and should be embraced by clinicians as an assistive tool. GRAPHICAL ABSTRACT
Background: The aim in this study was to investigate the localization of diabetic retinopathy features at the posterior pole. Methods: This study extracted diabetic retinopathy feature locations from 757 macula-centered 45-degree fundus photographs in the publicly available DDR dataset. Arteriole and venule locations were also extracted from the RITE (n = 35) and IOSTAR (n = 29) datasets. Images were normalized to collocate optic disc and macula positions, and feature positions were collated to generate a frequency distribution matrix. Sørensen–Dice coefficients were calculated to compare the location of different features. Results: Arterioles occurred in two main, distinct arcuate patterns. Venules showed a more diffuse distribution. Microaneurysms were diffusely located around the posterior pole. Hemorrhages and exudates occurred more frequently at the temporal aspect of the macula. Cotton wool spots occurred in a region approximating the radial peripapillary capillaries. Intraretinal microvascular abnormalities and neovascularization were seen throughout the posterior pole, with neovascularization at the disc (n = 65) being more common than neovascularization elsewhere (n = 46). Venous beading occurred primarily between the first and third bifurcations of the venules. Diabetic retinopathy overall was more frequent in the temporal aspect of the macula. The location of cotton wool spots and exudates showed moderate similarity (0.52) when all data were considered, reducing to low similarity (0.18) when areas of low frequency were removed. Conclusions: Diabetic retinopathy occurs throughout the posterior pole but is more frequent in the temporal aspect of the macula. Understanding the location of diabetic retinopathy features may help inform visual search strategies for diabetic retinopathy screening.
Objectives: This study investigated the visual search behaviour of optometrists and fellowship-trained ophthalmologists when screening for diabetic retinopathy in retinal photographs. Methods: Participants assessed and graded retinal photographs on a computer screen while a Gazepoint GP3 HD eye tracker recorded their eye movements. Areas of interest were derived from the raw data using Hidden Markov modelling. Fixation strings were extracted by matching raw fixation data to areas of interest and resolving ambiguities with graph search algorithms. Fixation strings were clustered using Affinity Propagation to determine search behaviours characteristic of the correct and incorrect response groups. Results: A total of 23 participants (15 optometrists and 8 ophthalmologists) completed the grading task, with each assessing 20 images. Visual search behaviour differed between correct and incorrect responses, with data suggesting correct responses followed a visual search strategy incorporating the optic disc, macula, superior arcade, and inferior arcade as areas of interest. Data from incorrect responses suggest search behaviour driven by saliency or a search pattern unrelated to anatomical landmarks. Referable diabetic retinopathy was correctly identified in 86% of cases. Grader accuracy was 64.8% with good inter-grader agreement (α = 0.818). Conclusions: Our study suggests that a structured visual search strategy is correlated with higher accuracy when assessing retinal photographs for diabetic retinopathy. Referable diabetic retinopathy is detected at high rates; however, there is disagreement between clinicians when determining a precise severity grade.
Purpose: To investigate how artificial intelligence (AI) systems detect referable diabetic retinopathy (DR) from retinal photographs by analysing heatmap patterns and determining their overlap with DR features. Methods: Fifty-four AI systems were developed using 27 backbone architectures, with each implemented as both binary-referable and multi-class grading models based on the International Clinical Diabetic Retinopathy (ICDR) grading scale. Models were trained on DDR, BRSET and Kaggle datasets. After training, each model analysed 749 images with DR feature annotations, with Grad-CAM, Integrated Gradients, LIME and SHAP heatmaps generated and compared to pixel-level annotations of microaneurysms, haemorrhages, exudates, cotton wool spots, venous beading, intraretinal microvascular abnormalities and neovascularisation. Results: All models achieved acceptable predictive performance (AUROC >0.8 for most architectures). Heatmap analysis revealed consistent attention to the central and temporal macular region. Exudates and cotton wool spots were highlighted most frequently by the heatmaps, with venous beading and neovascularisation at the disc showing poor overall coverage for binary referable classifiers. Models grading per the ICDR scale demonstrated higher coverage for all features. Substantial variability was observed between architectures, suggesting different feature detection capabilities. Interestingly, the heatmap analysis indicated that the models may be using different features to the ICDR grading scale definitions. Conclusion: AI models do not uniformly correlate with all DR features when detecting referable DR, limiting their predictive performance in unusual presentations. Heatmap aggregation analysis provides a scalable method for analysing model behaviour, allowing strengths and weaknesses to be assessed. These findings may help improve clinician’s trust and acceptance of AI.
Purpose: Measuring near vision provides clinicians with valuable insight into visual function. There is limited information on the accuracy of available reading charts frequently used in community practice. This study aimed to measure internationally available reading charts to determine how they compare to international standards. Methods: Commercially and device manufacturer-provided reading charts available in community optometry clinics were scanned at 600 dots per inch. Gaussian adaptive threshold was used to facilitate repeatable measurements. X- heights of letters were measured independently by three researchers. Contrast levels and line spacing were also estimated. Results for each chart were compared with ISO Standards. Intraclass correlation coefficient was used to assess intergrader agreement. Results: Of the 19 reading charts that were measured, only one chart (5.26%) had text sizes that were all within tolerance. There was high variability in size observed between charts. Twelve charts (63.2%) used serif fonts and seven used sans-serif (36.8%). Text on serif charts tended to be smaller than required (µ=−9.63%) compared to sans-serif (µ=+4.96%). All charts met the line spacing requirements and minimum required contrast level; however, some charts were printed on laminated or satin plastic which does not meet the standard of using a matte surface. There was high interrater agreement (ICC(2,1) = 1.00), indicating a highly repeatable measurement technique. Conclusion: This study found that the tested reading charts displayed significant variability in text size. Although some charts had more lines of text within size tolerances than others, none met all the requirements of the International Standard. Clinicians and researchers should take care when interpreting changes in near reading acuity when multiple charts have been used, especially as part of shared care models or when monitoring progressive vision changes. A free UC/UWA Reading Chart has been developed as a result of this study, which conforms to the ISO Standard. Key points: None of the measured reading charts met the requirements of the ISO 7921:2024 standard. There is high variability in text size between reading charts. A new chart, the UC/UWA Reading Chart, has been developed to conform to the ISO standards.