Artificial Intelligence in Ophthalmology

Introduction

Artificial intelligence(AI), often described as the pivotal technology of the fourth industrial revolution of mankind, refers to the theory and development of computer systems to be able to perform tasks normally requiring human intelligence. Deep learning (DL) is the latest subtype and most powerful AI tool. It uses a data driven approach to teach a computer to solve a task by itself, by giving inputs and their desired outputs as examples.It uses multiple layers of convolutional neural networks (CNNs),or Artificial neural networks (ANNs), mimicking the human brain, transforming inputs into an output.

Applications of AI in Ophthalmology

CNNs are highly suited for specialties like Ophthalmology which are driven by imaging modalities and hence research in this space has sparked tremendous global interest. Not surprisingly, the first US Food and Drug Administration (FDA)–approved AI diagnostic device was also in the field of ophthalmology for detection of more than mild diabetic retinopathy (mtmDR) by IDx-DR system. AI has achieved robust performance in the detection of DR, Retinopathy of Prematurity(ROP),Glaucoma, Age related macular degeneration(AMD) when applied to fundus images,OCT and visual fields.Scope in ophthalmology goes well beyond the posterior segment and includes refractive error prediction, detection of early keratoconus and strabismus, cataract grading, leukocoria detection, IOL power calculation, squamous cell neoplasia and even to an extent of predicting cardiovascular, stroke and Alzheimer’s risk from fundus images.(1)(2) While most algorithms provide diagnosis of a single disease from a single modality, future trend would be integrate data from multiple sources and modalities that might provide a more accurate diagnosis of several disease with their severity at one time.

A vast majority of the AI work has revolved around the leading causes of preventable blindness in the world,to harness AI’s potential for early detection. Abramoff etal, in their pivotal clinical trial reported that their DL model was able to achieve a sensitivity and specificity of 87.2% and 90.7% respectively in the detection of mtmDR.(3)Several other groups Tingetal,(1)Gulshanetal,(4)Natarajanetal,(5) Rajalakshmi etal(6)and Sosale etal,(7)have been instrumental in the research on applying AI for DR screening. In the detection of AMD, of note is the work of Grassmann et al with a sensitivity of 84.2% in the detection of any AMD, (8) Ting et al with a performance of 93.2% sensitivity of detecting referable AMD,(9) and Burlina etal with an AUCof0.94-0.6 using fundus images.(2)(10)Additionally, research groups have explored possibilities of using AI in detecting cases of AMD using OCT images as well.(1) (2) Most recently, Brown et al have reported impressive results of DL system that can detect plus disease,a treating requiring form of severe ROP with an AUCof0.98.(11)The iROP tool developed by the group also seems promising as it allows for objective monitoring of plus disease by producing a severity score. While glaucoma is a far more challenging diagnosis to make, groups like Tingetaland Li et al have developed algorithms to detect glaucoma-like discs from fundus image with acceptable accuracy. (1) AI models have also been attempted to analyse visual field data to predict future loss and monitor progression. (12) Researchers have even explored the possibility of detecting retinal nerve fibre layer damage in glaucoma using OCT scans. (1) Following the successful development of AI algorithms using fundus images, De Fauw et al at DeepMind in 2018 applied DL to OCT, developing a model that could detect 50 ophthalmic pathologies, by automatically segmenting tissue layers and achieving a diagnostic performance exceeding those of expert human graders.(13)

Despite several groups having demonstrated promising performance of AI,there remains a significant deficit in the prospective validation of these algorithms. Following the precedent set by IDx and now EyeArt which recently received FDA approval,AI systems which are classified as “Software for Medical

Device” will require to demonstrate acceptable performance in real world prospective trials for

regulatory approval.

Innovation in AI: Remidio in focus

Remidio Innovative Solutions Pvt Ltd. has taken an approach to develop fundus imaging and AI screening as a single product. Fundus on phone(FOP)(Figure1)is a portable,affordable,smartphone- based fundus camera with patented optics that has been validated against high-end traditional desktop cameras for image quality.(14) The differentiating feature is that FOP comes integrated with an offline, on the edge Medios AI that can generate a report for referable diabetic retinopathy (RDR) in less than 10 seconds. (6) Thus, it eliminates challenges associated with cloud-based inferencing by drastically reducing the turn-around time for reporting as well the need for internet connectivity and heavy infrastructure support.

Figure 1: Remidio FOP NM-10, Non-mydriatic, Smartphone-based Fundus camera with integrated Medios AI

Medios AI for screening of RDR has been prospectively validated in clinical trials both at tertiary care diabetes centres as well as during opportunistic screening in a population setting. (7)(5) The performance has been above the FDA mandated superiority end points. With a sensitivity and specificity of 100.0% and 88.4%, respectively, for RDR (Natarajan et al)(5) and 93% sensitivity for RDR and 95.5% specificity for any DR (Sosale et al)(7), this AI algorithm is a landmark in the adaptation of AI models for community-based DR screening. The algorithm also uses heat maps to highlight probable lesions in grades of referable diabetic retinopathy enabling the operator to educate the patient on the need for immediate specialist consult (Figure2).

Figure 2: Sample report generated by Medios AI with activation maps for different stages of Diabetic Retinopathy

Moving a step ahead, another layer has been added to this algorithm to enable further staging and triaging. On the images that have been flagged as RDR positive, another algorithm detects the presence of Sight-threatening DR(STDR; defined as either Severe NPDR or Proliferative DR and/orthe presence of Diabetic macular edemaDME).(15)This algorithm has also been validated in a prospective trial with sensitivity of 95.59% and specificity of 94.6% and is soon to be deployed on the device as well. (9) Remidio is also working on several other screening algorithms for glaucoma, AMD, ROP and cataract which are in different stages of development and validation.This approach of having a single device-portable fundus camera along with integrated offline AI,can potentially make a paradigm shift in the approach for screening as providers could get to the population to screen at-risk individuals rather than relying on patients to come to the point of care to be screened.

Conclusion

Recent advances in AI present an exciting opportunity to improve and revolutionize ophthalmic care. The scope of AI has expanded to several realms of ophthalmology and can potentially reduce the barriers of healthcare access and screening, thus reducing the global burden of avoidable blindness. AI can not only serve as a triaging tool but also aid in diagnosis for ophthalmologists enabling overall efficient use of resources. Whilst there are several potential benefits of introducing AI systems, there are challenges that can be encountered that can act as a hindrance to scale this technology- requirement of diverse and adequate datasets, understanding the ‘black box’ of the model, technical challenges for offline deployment, robust real world prospective trials, medico-legal and ethical considerations, regulatory hurdles and barriers of patient acceptability. We must remain cognisant of the potential unintended consequences and ensure that robust AI algorithms are developed with global community in mind. Further work is essential for successful clinical adoption of AI and this will ultimately be achieved if the implementation is safe, evidence-based, and meaningful in improving patient outcomes.

REFERENCES:

 

  1. Ting DSW, Pasquale LR, Peng L, Campbell JP, Lee AY, Raman R, et al. Artificial intelligence and deep learning in ophthalmology. Br J Ophthalmol. 2019 Feb1;103(2):167–75.
  2. Dutt S, Sivaraman A, Savoy F, Rajalakshmi R. Insights into the growing popularity of artificial intelligence in ophthalmology. Indian J Ophthalmol. 2020 Jul1;68(7):1339.
  3. Abràmoff MD, Lavin PT, Birch M, Shah N, Folk JC. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. Npj Digit Med. 2018 Aug28;1(1):1–8.
  4. Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Narayanaswamy A, et al. Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. JAMA. 2016 Dec13;316(22):2402.
  5. Natarajan S, Jain A, Krishnan R, Rogye A, Sivaprasad S. Diagnostic Accuracy of Community-Based Diabetic Retinopathy Screening With an Offline Artificial Intelligence System on a Smartphone. JAMA Ophthalmol. 2019 Oct1;137(10):1182–8.
  6. Rajalakshmi R, Subashini R, Anjana RM, Mohan V. Automated diabetic retinopathy detection in smartphone-based fundus photography using artificial intelligence.Eye. 2018Jun;32(6):1138–44.
  7. Sosale B, Aravind SR, Murthy H, Narayana S, Sharma U, Gowda SGV, et al. Simple, Mobile-based Artificial Intelligence Algo r ithm in the detection of Diabetic Retinopathy (SMART) study. BMJ Open Diabetes Res Care. 2020Jan;8(1):e000892.
  8. Grassmann F, Mengelkamp J, Brandl C, Harsch S, Zimmermann ME, Linkohr B, et al. A Deep Learning Algorithm for Prediction of Age-Related Eye Disease Study Severity Scale for Age- Related Macular Degeneration from Color Fundus Photography. Ophthalmology. 2018 Sep 1;125(9):1410–20.
  9. Ting DSW, Cheung CY-L, Lim G, Tan GSW, Quang ND, Gan A, et al. Development and Validation of a Deep Learning System for Diabetic Retinopathy and Related Eye Diseases Using Retinal Images From Multiethnic Populations With Diabetes. JAMA. 2017 Dec12;318(22):2211–23.
  10. Burlina PM, Joshi N, Pekala M, Pacheco KD, Freund DE, Bressler NM. Automated Grading of Age- Related Macular Degeneration From Color Fundus Images Using Deep Convolutional Neural Networks. JAMA Ophthalmol. 2017 Nov1;135(11):1170–6.
  11. Brown JM, Campbell JP, Beers A, Chang K, Ostmo S, Chan RVP, et al. Automated Diagnosis of Plus Disease in Retinopathy of Prematurity Using Deep Convolutional Neural Networks. JAMA Ophthalmol. 2018 Jul1;136(7):803–10.
  12. Wang M, Shen LQ, Pasquale LR, Boland MV, Wellik SR, Moraes CGD, et al. Artificial Intelligence Classification of Central Visual Field Patterns in Glaucoma. Ophthalmology. 2020 Jun 1;127(6):731–8.
  13. Fauw JD, Ledsam JR, Romera-Paredes B, Nikolov S, Tomasev N, Blackwell S, et al. Clinically applicable deep learning for diagnosis and referral in retinal disease. Nat Med. 2018 Sep;24(9):1342–50.
  14. Sengupta S, Sindal MD, Baskaran P, Pan U, Venkatesh R. Sensitivity and Specificity of Smartphone-Based Retinal Imaging for Diabetic Retinopathy: A Comparative Study. Ophthalmol Retina. 2019 Feb1;3(2):146–53.
  15. Parthasarathy DR, Savoy F, Sosale B, Sosale A, Murthy H, Narayana S, et al. Performance of screening algorithms for Referable Diabetic Retinopathy (RDR) and Sight Threatening Diabetic Retinopathy (STDR) on a non-mydriatic portable smartphone-based fundus camera. Invest Ophthalmol Vis Sci. 2020 Jun10;61(7):5319–5319.

 

By,
Dr. Divya P Rao, MS, FLVP (Glaucoma),
Medical Director,
Remidio Innovative Solutions

Leave a Reply

Your email address will not be published. Required fields are marked *