You are here: » Vision Research » The Young Researchers View » The Young Researcher's View: Reena Chopra (Q4-2020)

The Research of Reena Chopra


I am an Optometrist at Moorfields Eye Hospital, London, United Kingdom, a PhD student at UCL Institute of Ophthalmology, and a Clinician Scientist at Google Health. During my PhD I completed an internship with DeepMind.

Much of my PhD work focuses on optical coherence tomography (OCT) imaging and the ability to use this modality to automate aspects of the eye examination. This includes hardware advances as well as the application of artificial intelligence to read these complex images.

Automating OCT device operation

Optical coherence tomography (OCT) devices are becoming ubiquitous in eye clinics worldwide. Much of this uptake is attributed to its ability to non-invasively capture micron-resolution images, enabling objective and quantitative data to be obtained from ocular structures - aiding the diagnosis and monitoring of eye disease. Although reasonably quick and safe to perform, the costs involved with operating OCT devices are not trivial. Commercial devices are expensive to purchase, and the costs of associated labour to capture the OCT scans are even greater. The requirement for OCT and other imaging is placing increasing demand on ophthalmology clinics, contributing to fragmented patient pathways and often extended waiting times. Furthermore, OCT devices are large, require significant floor space, and have limited portability to use outside of typical clinical settings.

My PhD work focuses on a new form of eye examination that has been developed by Envision Diagnostics, Inc. (El Segundo, CA, USA) termed Binocular Optical Coherence Tomography1, which aims to address the shortcomings of the current OCT instruments and add many unique capabilities. Unlike other OCT devices, the binocular OCT prototype consists of two oculars, which enables a pair of eyes to be imaged simultaneously. The oculars align to the patients’ eyes automatically, and thus an operator isn’t required to move the device from one eye to the other, and instead can be self-operated by the user. The binocular OCT also utilises a tunable swept-source laser system with adjustable optics that can switch from anterior eye imaging to lens imaging, to vitreous imaging, and to posterior pole imaging, permitting whole-eye OCT without the need for additional attachments. Furthermore, the device is equipped with ‘smart technology’, offering more advanced display, input and computing capabilities than conventional OCT. A speaker system is used to deliver audio instructions to guide the automated examination. In addition, the binocularity aspect can be exploited so that OCT imaging can be used for novel applications such as diagnostic functional tests including pupillometry, strabismus measurement and ocular motility. Traditionally these tests have been subjective and required significant clinical expertise to interpret.

Structured, patient-centred, usability testing is essential to the design, clinical validation, regulatory approval, and widespread implementation of all new medical devices 2. This is particularly the case for a putative binocular OCT system – an instrument intended for automated use in visually impaired, often elderly, populations. In a usability study, a cohort of first time users (45 participants with chronic eye conditions and 15 participants without eye conditions) were able to complete a comprehensive suite of tests using the binocular OCT prototype without any previous training or assistance during the examination3. Only a small proportion of examinations generated ungradable data for diagnostic tests such as visual acuity and visual fields testing due to reduced sensitivity of the voice recognition system.

We have performed early-stage diagnostic studies to assess the agreement, repeatability, and reproducibility of individual diagnostic features of the device across a range of conditions, by comparing their results with reference standards. Initial studies have shown that the device was able to correctly identify the direction of strabismus including both horizontal and vertical elements (Figure 1)4. Similarly promising results have been shown for binocular OCT pupillometry4,5. The device was found to have excellent test-retest reliability for pupil parameters such as maximum and minimum pupil diameter, and anisocoria. Although the diagnostic accuracy for detection of relative afferent pupillary defect was inferior to another automated pupillometer (Konan RAPDx), OCT-derived measurements show promise in detecting these, often subtle, abnormalities (Figure 2). The next steps include further testing, and evaluating the health economic implications of implementing an automated device in an eye clinic.

Figure 1. The right eye is the fixating eye (A), and the left eye is the strabismic eye (B). The angle of the deviation is calculated by measuring the tilt of the eye with respect to the fixating eye. The pupil margins are used as landmarks to measure tilt. This pair of images indicates a left esotropia measuring an angle of 10.99o. N indicates nasal; T, temporal.
Figure 2. a) Resting diameters pre-stimulus; b) Flash presented to the left eye, constriction of both pupils observed; c) Both pupils dilated to their resting diameter; d) Flash presented to the right eye. Constriction amplitude of both eyes was less than that observed when the flash was presented to the left eye.

Automating OCT image interpretation

Ocular images are information-rich, and conceal a surprising amount of information that even the most experienced clinicians cannot detect. Artificial intelligence (AI), more specifically deep learning models, can identify an individual’s biological sex, cardiovascular risk factors6, and refractive error7 from a fundus photograph alone. So far within ophthalmology, AI has been primed to predict events at the time that the image was taken. In a collaboration between Moorfields Eye Hospital and DeepMind, we were able to successfully use AI to classify and triage retinal eye diseases from OCT images8. Here, the model did not miss a single urgent case. Can we use similar deep learning methods to develop an early warning system that uses images to predict the future onset of disease?

In Europe, it is estimated that 25% of the adults over age 60 have some form of age-related macular degeneration (AMD)9. Although there is little treatment other than lifestyle changes for dry AMD, sight-preserving treatment is available for ‘wet’ form of the disease. In a collaboration between Moorfields Eye Hospital, DeepMind, and Google Health, we trained an AI-system to predict whether a patient with wet AMD in one eye will develop the condition in their second eye10. Within two years of diagnosis, 20% of patients develop wet AMD in their second (fellow) eye11,12 – often the better-seeing eye that individuals are reliant upon day-to-day. The period before the development of wet AMD may be a critical window for planning of follow-up intervals, and even for potential administration of preventative treatments that are currently in clinical trials.

For this work, we curated an anonymised dataset from Moorfields of patients receiving treatment for wet AMD in one eye. We used OCT images of the fellow eye at each treatment visit. These images were input into a two-level AI system. The first level segmented the OCT image into several different morphological features, such as drusen, retinal pigment epithelium, and hyperreflective foci. The second level input both the raw OCT and the segmented image from the first level. The output of the system provided an estimate of a patient’s risk of progressing to wet AMD within the ensuing six months.

As this task is not routinely performed in practice, clinician performance for future prediction was unknown. We recruited 3 ophthalmologists and 3 optometrists specialising in Medical Retina to make similar predictions, firstly using the OCT alone (akin to the AI-system), and secondly using the OCT, fundus photograph, and all available clinical information (replicating a real-life clinical scenario). We found that clinicians could perform this task but with substantial variability in sensitivity and specificity among them. The AI system had significantly better sensitivity and specificity than five out of six specialists.

One significant benefit of an AI system is that it can be tuned to the clinical application. In the paper, we discuss 2 operating points based on specific sensitivity and specificity thresholds. This impacts the number of false positives. If one was considering an invasive preventative treatment, the operating point could be tuned to reduce the false positive rate. Whereas, if the system was simply informing follow-up intervals, a more liberal operating point might be chosen where the sensitivity is greater but also allowing more false positives. Another benefit of this system is that it produces segmentation maps as an intermediate step. These maps enable clinicians to visualise the longitudinal changes in macular morphology, and may improve the understanding of disease progression (Figure 3).

These findings demonstrate the potential for AI to help improve the understanding of disease progression and predict the future risk of patients developing sight-threatening conditions. In the future, both automation and artificial intelligence are likely to re-engineer medicine to cope with some of the pressures of increasing demand and strained healthcare systems. Applied to OCT imaging, this could usher in a new era of comprehensive eye care that has the potential to transform the practice of ophthalmology.

Figure 3. Example of a correct prediction by the AI system. a) The OCT at each visit is segmented using the deep learning model. In the months leading to conversion, we observed an increasing presence of subretinal hyperreflective material (SHRM) and fibrovascular pigment epithelium detachment (fibro PED). The eye converts to wet AMD at 10.5 months and starts treatment 2 months later. b) Prediction of the AI system for conversion to wet AMD within 6 months. At the liberal operating point (yellow dotted lines) it correctly predicted conversion within 6 months for all three scans within the actual 6-month window before conversion (gray box).


  1. Walsh, A. C. Binocular optical coherence tomography. Ophthalmic Surg. Lasers Imaging 42 Suppl, S95–S105 (2011).
  2. Michael E. Wiklund, P. E., Kendler, J. & Strochlic, A. Y. Usability Testing of Medical Devices, Second Edition. (CRC Press, 2015).
  3. Chopra, R., Mulholland, P. J., Dubis, A. M., Anderson, R. S. & Keane, P. A. Human Factor and Usability Testing of a Binocular Optical Coherence Tomography System. Transl. Vis. Sci. Technol. 6, 16 (2017).
  4. Chopra, R., Mulholland, P. J., Tailor, V. K., Anderson, R. S. & Keane, P. A. Use of a Binocular Optical Coherence Tomography System to Evaluate Strabismus in Primary Position. JAMA Ophthalmol. 136, 811–817 (2018).
  5. Chopra, R. et al. Automated Pupillometry Using a Prototype Binocular Optical Coherence Tomography System. Am. J. Ophthalmol. 214, 21–31 (2020).
  6. Poplin, R. et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng 2, 158–164 (2018).
  7. Varadarajan, A. V. et al. Deep Learning for Predicting Refractive Error From Retinal Fundus Images. Invest. Ophthalmol. Vis. Sci. 59, 2861–2868 (2018).
  8. De Fauw, J. et al. Clinically applicable deep learning for diagnosis and referral in retinal disease. Nat. Med. 24, 1342–1350 (2018).
  9. Li, J. Q. et al. Prevalence and incidence of age-related macular degeneration in Europe: a systematic review and meta-analysis. Br. J. Ophthalmol. 104, 1077–1084 (2020).
  10. Yim, J. et al. Predicting conversion to wet age-related macular degeneration using deep learning. Nat. Med. (2020) doi:10.1038/s41591-020-0867-7.
  11. Zarranz-Ventura, J. et al. The neovascular age-related macular degeneration database: report 2: incidence, management, and visual outcomes of second treated eyes. Ophthalmology 121, 1966–1975 (2014).
  12. Fasler, K. et al. Moorfields AMD database report 2: fellow eye involvement with neovascular age-related macular degeneration. Br. J. Ophthalmol. 104, 684–690 (2020).

Reena Chopra

NIHR Biomedical Research Centre for Ophthalmology, Moorfields Eye Hospital NHS Foundation Trust and UCL Institute of Ophthalmology,
London, United Kingdom

E-mail: reena.chopra1[at]