Deep Learning Automated Diagnosis and Grading of Cataracts using Colour Fundus Images: The Fundus Cataract-AI Project

Abstract

Purpose : Current deep learning systems for automated cataract diagnosis and grading focus on the use of slit lamp images. Acquisition of such images is however difficult in primary care owing the technical skills required to operate a slit lamp biomicroscope. A fundus camera is much more widely available and accessible tool in primary care. The Fundus Cataract-AI project aims to automate cataract diagnosis and grading using deep learning techniques applied to a single standard macula-centric fundus photo.

Methods : A dataset was utilized, comprising 11544 fundus images from Chinese patients, aged 50 and above, from a cross-sectional random population-based study. 909 images of patients who underwent cataract surgery previously were removed from the dataset. Another 3674 images were excluded from the dataset due to poor image quality. The subjects underwent comprehensive eye assessment by an Ophthalmologist including presenting visual acuity (PVA), best-corrected visual acuity (BCVA), automatic and subjective refraction, IOP, slit-lamp examination of the anterior segment and funduscopic examination of the posterior segment. Fundus photography was performed using the Daytona Ultra-widefield Fundus Camera (Optos, Inc., Dunfermline, Scotland, UK). The images were categorized into three classes: ‘normal’, 'early cataract', 'visually significant cataract' based on face-to-face assessment by the Ophthalmologist and BCVA. A ResNet152 convolutional neural network, pretrained and fine-tuned via transfer learning, was employed for classification. The dataset was split into training (60%), validation (20%) and testing (20%) sets. Main outcome measure was area under the curve (AUC) with accuracy, sensitivity and specificity.

Results : 6961 images were used for the dataset in total. The the distribution of ‘normal’, ‘early cataract’ and ‘visually significant cataract’ was 31.7%, 45.9% and 13.0% respectively. The preliminary model demonstrated an overall performance accuracy of 75%. Review of heat maps show the optic disc and macula area highlighted as areas of interest by the algorithm.

Conclusions : Machine learning may be able to accurately assist in the diagnosis and grading of cataracts using fundus photos alone. Further refinement is needed to improve performance accuracy of this algorithm.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Confusion matrix showing performance of algorithm in determining cataract presence and severity

 

Heat Map