Article Text

Download PDFPDF
93 A machine learning approach to classification of uveitis in children using anterior segment OCT images of the eye
  1. Usmaan Bhatti1,
  2. Saira Akbarali2,
  3. Ameenat L Solebo2,
  4. William Bryant2
  1. 1Imperial College London
  2. 2Great Ormond Street Hospital, London, UK


Context Grading of uveitis is currently assessed by slit-lamp examination based on the SUN scale, which is subjective and open to interobserver variation. ASOCT imaging has been proposed as an alternative, but manual interpretation of large volumes of scans is time-consuming. The proliferation of OCT machines across the country will serve to increase the burden on clinicians. Automation of OCT interpretation using deep learning has recently been shown to be effective for assessing retinal pathologies.

Aims This project aims to partially automate the interpretation of ASOCT images by providing automated white cell counts by utilising image processing techniques and two deep neural networks to both assess the viability of automating white cell counts in ASOCT images and find the optimal system to obtain reliable counts.

Methods ASOCT images obtained from children with and without uveitis were analysed by an image processing algorithm, identifying areas of potential white cells, which were labelled by an ophthalmologist and an optician, of which 30 × 30 pixel images were analysed by two deep learning binary classifiers; a Multi-Layer Perceptron (MLP) and a Convolutional Neural Network (CNN).

Results In total, 50 ASOCT images were obtained from 26 children, with 620 potential white cells identified (10% positively labelled). Mean accuracy MLP/CNN: 0.937/0.952, t-test p-value=0.01535. Mean F1-Score MLP/CNN: 0.644/0.778, t-test p-value=0.00061. Mean AUC-PR curve MLP/CNN: 0.725/0.814. Correct classification proportion of positive label MLP/CNN: 0.54/0.80, McNemar’s test p-value=0.00615

Conclusion The CNN outperformed the MLP in all measures, especially in identifying the positive class. Correlations with manual counting and clinical grade, expansion of training data, and functionality to measure anterior segment flare should all be considered to assess and improve clinical utility. However, automated counting shows promise as a viable means of measuring white cells in ASOCT images.

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.