Artificialintelligence(AI),oftendescribedasthepivotaltechnologyofthefourthindustrialrevolution 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 theirdesiredoutputsasexamples.Itusesmultiplelayersofconvolutionalneuralnetworks(CNNs),or Artificial neural networks (ANNs), mimicking the human brain, transforming inputs into anoutput.
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 hasachievedrobustperformanceinthedetectionofDR,RetinopathyofPrematurity(ROP),Glaucoma, Age-relatedmaculardegeneration(AMD)whenappliedtofundusimages,OCTandvisualfields.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’sriskfromfundusimages.(1)(2)Whilemostalgorithmsprovidediagnosisofasingledisease from a single modality, future trend would be integrate data from multiple sources and modalities thatmightprovideamoreaccuratediagnosisofseveraldiseasewiththeirseverityatonetime.
AvastmajorityoftheAIworkhasrevolvedaroundtheleadingcausesofpreventableblindnessinthe world,toharnessAI’spotentialforearlydetection.Abramoffetal,intheirpivotalclinicaltrialreported that their DL model was able to achieve a sensitivity and specificity of 87.2% and 90.7% respectively inthedetectionofmtmDR.(3)SeveralothergroupsTingetal,(1)Gulshanetal,(4)Natarajanetal,(5) Rajalakshmietal(6)andSosaleetal,(7)havebeeninstrumentalintheresearchonapplyingAIforDR 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 referableAMD,(9)andBurlinaetalwithanAUCof0.94-0.6usingfundusimages.(2)(10)Additionally, researchgroupshaveexploredpossibilitiesofusingAIindetectingcasesofAMDusingOCTimagesas well.(1) (2) Most recently, Brown et al have reported impressive results of DL system that can detect plusdisease,atreatingrequiringformofsevereROPwithanAUCof0.98.(11)TheiROPtooldeveloped by the group also seems promising as it allows for objective monitoring of plus disease by producing aseverityscore.Whileglaucomaisafarmorechallengingdiagnosistomake,groupslikeTingetaland 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 humangraders.(13)
DespiteseveralgroupshavingdemonstratedpromisingperformanceofAI,thereremainsasignificant deficit in the prospective validation of these algorithms. Following the precedent set by IDx and now EyeArtwhichrecentlyreceivedFDAapproval,AIsystemswhichareclassifiedas“SoftwareforMedical
Device” will require to demonstrate acceptable performance in real world prospective trials for
Remidio Innovative Solutions Pvt Ltd. has taken an approach to develop fundus imaging and AI screeningasasingleproduct.Fundusonphone(FOP)(Figure1)isaportable,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 infrastructuresupport.
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 AImodelsforcommunity-basedDRscreening.Thealgorithmalsousesheatmapstohighlightprobable lesionsingradesofreferablediabeticretinopathyenablingtheoperatortoeducatethepatientonthe 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 presenceofSight-threateningDR(STDR;definedaseitherSevereNPDRorProliferativeDRand/orthe presenceofDiabeticmacularedemaDME).(15)Thisalgorithmhasalsobeenvalidatedinaprospective 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 cataractwhichareindifferentstagesofdevelopmentandvalidation.Thisapproachofhavingasingle device-portablefunduscameraalongwithintegratedofflineAI,canpotentiallymakeaparadigmshift 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 bescreened.
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,regulatoryhurdlesandbarriersofpatientacceptability.Wemustremaincognisantof 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 patientoutcomes.
Dr. Divya P Rao, MS, FLVP (Glaucoma)
Remidio Innovative Solutions