Automated Diagnosis And Cataract Grading In Chinese Patients Presenting With Blurring Of Vision For Three Or More Months Using Deep Learning Techniques Applied To Colour Retina Photographs

The aim of the study was to train, tune and validate a model for a DLS-assisted clinical decision aid for triage of ophthalmic referrals and cataract grading for non-diabetic Chinese patients, aged 50 and over, with a presenting complaint of chronic blurring of vision for three or more months.

Deep learning system-generated algorithms have the capacity to support clinical decision making in medical practice, particularly in fields where the process is subjective and conducted with limited information at hand. One such field is in triage of new patient referrals for ophthalmic care at government specialist outpatient clinics in Hong Kong.

First phase of study consisted of training, tuning, and internal validation of an automated DLS to triage ophthalmic referrals and grade cataracts. A ResNet-50 convolutional neural network was developed to create a binary classification model using defined clinical criteria to split labelled retinal photographs into ‘early’ and ‘routine’ referrals. Neural networks ResNet-50, DenseNet-121, and Vision Transformer b16 were used to develop three-class models, classifying labelled retinal photos as ‘normal’, ‘early’, or ‘visually significant’ cataracts based on predefined criteria. External validation evaluated performance using the area under the receiver operating characteristic curve, sensitivity, specificity, and Cohen’s kappa statistic.

The binary classification model achieved an AUC of 0.998 in internal validation, with 100% sensitivity and 98.3% specificity indicating high performance. In comparison to the binary classification model, current triage practice performed relatively poorly, on validation testing, scoring an AUC, sensitivity and specificity of 0.656, 42.1% and 89.1% respectively. Across the different CNNs, the three-class classification model for cataract grading achieved weighted average AUCs between 0.807 and 0.83 and macro average AUCs between 0.837 and 0.868. Agreement level with ophthalmologist assessment was only moderate for the binary classification model but substantial for the three-class classification models.

Our automated DLS model is an effective and accurate clinical decision aide in triaging non-urgent new case referrals to ophthalmology and in grading of cataracts based on the data from a single macula-centric retina photograph taken without mydriasis.