medios– an offline, smartphone-based artificial ... · medios– an offline, smartphone-based...

Post on 28-Jul-2020

5 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

© 2020 Indian Journal of Ophthalmology | Published by Wolters Kluwer - Medknow

Original Article

Medios– An offline, smartphone-based artificial intelligence algorithm for the diagnosis of diabetic retinopathy

Bhavana Sosale, Aravind R Sosale, Hemanth Murthy1, Sabyasachi Sengupta2, Muralidhar Naveenam1

Access this article onlineWebsite: www.ijo.inDOI: 10.4103/ijo.IJO_1203_19PMID: *****

Quick Response Code:

Purpose: Anobservationalstudy toassessthesensitivityandspecificityoftheMediossmartphone-basedoffline deep learning artificial intelligence (AI) software to detect diabetic retinopathy (DR) comparedwith the image diagnosis of ophthalmologists. Methods: Patients attending the outpatient services of atertiarycenterfordiabetescareunderwent3-fielddilatedretinalimagingusingtheRemidioNMFOP10.Twofellowship-trainedvitreoretinalspecialistsseparatelygradedanonymized imagesandapatient-leveldiagnosiswasreachedbasedongradingoftheworseeye.TheimagesweresubjectedtoofflinegradingusingtheMediosintegratedAI-basedsoftwareonthesamesmartphoneusedtoacquireimages.ThesensitivityandspecificityoftheAIindetectingreferableDR(moderatenon-proliferativeDR(NPDR)orworsedisease)was compared to the gold standard diagnosis of the retina specialists. Results: Results include analysisof imagesfrom297patientsofwhich176(59.2%)hadnoDR,35(11.7%)hadmildNPDR,41(13.8%)hadmoderateNPDR,and33(11.1%)hadsevereNPDR.Inaddition,12(4%)patientshadPDRand36(20.4%)hadmacularedema.SensitivityandspecificityoftheAIindetectingreferableDRwas98.84%(95%confidenceinterval [CI],97.62–100%)and86.73%(95%CI,82.87–90.59%),respectively.Theareaunderthecurvewas0.92.Thesensitivityforvision-threateningDR(VTDR)was100%.Conclusion: TheAI-basedsoftwarehadhigh sensitivity and specificity indetecting referableDR. Integrationwith the smartphone-based funduscamerawithofflineimagegradinghasthepotentialforwidespreadapplicationsinresource-poorsettings.

Key words: Artificialintelligence,deeplearning,diabeticretinopathy

Department of Diabetology, DiaconHospital, 1Department of Vitreo-RetinalSurgery,RetinaInstituteofKarnataka,2Department of Vitreo-Retinal Surgery, Future Vision Eye Care,Mumbai,Maharashtra, India

Correspondence to: Dr. Bhavana Sosale, 360,DiaconHospital,19thMail,1stBlock,Rajajinagar,Bengaluru-560010,Karnataka,India.E-mail:bhavanasosale@gmail.com

Received:26-Jun-2019 Revision: 16-Oct-2019Accepted:18-Nov-2019 Published:20-Jan-2020

AroundfivemillionIndianshavevision-threateningdiabeticretinopathy (VTDR).[1,2] Smartphone-based fundus camerasandevolutionofartificialintelligence(AI)canmakescreeningscalable.[3-8]HighcomputationalpowerandInternetaccess—aprerequisiteforcloud-basedAI—isoftenlackingindevelopingcountries.MediosTechnologies,SingaporetoourknowledgeisthefirstcompanytodevelopanofflineAIalgorithmtoaddressthisobstacle.

Our aim is to evaluate the performance of an offlineAI-basedsoftware(MediosTechnologies,Singapore)loadedon a smartphone-based fundus camera in thedetectionofdiabeticretinopathy(DR)comparedwiththeimagediagnosisof ophthalmologists.

MethodsThestudywasapprovedbytheinstitutionalethicscommitteeandcarriedoutasperthedeclarationofHelsinki.Patientswhoconsentedtohavetheireyedilatedandphotographedduringroutinecarewereincluded.

Thiswas a cross-sectional observational study of 304diabeticpatientsattendingtheoutpatientdepartment(OPD)of auniversity-recognized tertiary center fordiabetes care

andresearchinBangalore,IndiaduringthemonthofOctober2018.All subjects, above 18 years of age,with type 1 or2diabetesorsecondarydiabeteswereinvitedtoparticipate.Eyeswithsignificantmediaopacity,suchascornealopacityoradvancedcataracts,thatprecludedretinalimagingwereexcluded.

Atthetimeofthepatient’shospitalvisit,routineclinicalcareprocedureslikethecollectionofdemographicdata,medicalhistory,vitalmeasurements,anthropometricmeasurements,andgeneralphysicalexamswerecarriedout.

Retinal image acquisitionAdropof1%tropicamidesolutionwasusedtodilatethepupilstoaminimumsizeof5mm.Retinalimageswerecapturedusingthe smartphone-based“RemidioNonMydriaticFundusonPhoneCamera(NMFOP10)”(RemidioInnovativeSolutionsPvt.Ltd.,Bangalore, India)bya trained technician.[3] Three fieldsofview(FOVs)werecapturedfromeacheye—posteriorpole(maculacentered),nasalfield,andsuperotemporalfield.ThetechnicianwastrainedtorecognizecharacteristicsofanexcellentimageandurgedtocapturemorethanoneimageperFOVifrequiredtoobtainexcellentimages.

Cite this article as: Sosale B, Sosale AR, Murthy H, Sengupta S, Naveenam M. Medios– An offline, smartphone-based artificial intelligence algorithm for the diagnosis of diabetic retinopathy. Indian J Ophthalmol 2020;68:391-5.

This is an open access journal, and articles are distributed under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 License, which allows others to remix, tweak, and build upon the work non-commercially, as long as appropriate credit is given and the new creations are licensed under the identical terms.

For reprints contact: reprints@medknow.com

[Downloaded free from http://www.ijo.in on Monday, January 20, 2020, IP: 106.206.50.176]

392 Indian Journal of Ophthalmology Volume68Issue2

The Remidio NM FOP 10 device uses an iPhone6smartphone’scameratocaptureimagesoftheretinaeitherby using an infrared light emittingdiode (IRLED) basedliveview(nonmydriaticmode)orusingawarmwhiteLEDliveview (themydriaticmode).Asper theApple’sofficialwebsite,theiPhone6specificationsincludeacamerawitha8MP(2448×3264pixels)resolution,ascreendisplayresolutionof750×1334pixels,anda1.4GHzCycloneprocessorpairedwith1GBofRAM.Duringthecourseofthisstudy,thephoneusedaniOS10mobileoperatingsystemandcamepreloadedwiththeintegratedRemidioNMFOP10appwiththeMediosAI.Themanufacturer stated resolutionof theRemidioNMFOPcamerawasaminimumof80linepairs/mm,conformingtotherequirementsofISO10940standard.

Image grading by a vitreoretinal specialistThede-identified (anonymized) imageswith the subject IDwere uploaded online from theNMFOP 10device to anAmazonWebServices(AWS)hostedcloudserviceprovidedbythemanufacturer.Theimageswereaccessedfromthecloudby two fellowship-trainedvitreoretinal surgeonswithmorethan20yearsof experience in treatingDR.Both the retinalsurgeons,gradingandadjudicatingtheimages,wereaffiliatedtoadifferenthospital;hence,theyremainedunbiased,beingmaskedtotheclinicalandAIdiagnosis.

The retinal surgeons individually graded the set of three retinal photographs fromeveryeyeusing the InternationalClinicalDiabeticRetinopathySeverityScaleSystem(ICDRS).[9] Images weregradedasnoDR,mildnonproliferativeDR(mildNPDR),moderate nonproliferativeDR (moderateNPDR), severenonproliferativeDR(severeNPDR),andproliferateDR(PDR).Thediagnosisofdiabeticmacularedema(DME)wasrecordedaspresentorabsent.Theeyewith themore severe stageofretinopathywas considered as thefinaldiagnosis for thatpatient,incaseswhereeacheyehadadifferentstageofdiseaseseverity.Patientswhoseimageswereconsideredungradablebytheretinaspecialistswereexcludedfromthefinalanalysis.Wheneverthetwogradersdifferedonthediagnosis,aconsensuswas reachedby revisiting the images,discussing them,andreachingamutualagreement.Theadjudicatedpatientdiagnosisobtainedfromtheretinaspecialistswasconsideredasthegoldstandardforcomparisons.

Theclinicaldiagnosiswasnotconsideredasthereferencestandard in this study. The retina specialists affiliated tothe tertiary diabetes hospital conducting the studyweremany—each covering the outpatient clinic on rotation ondifferentdaysoftheweek.Thismadeadjudicationofclinicaldiagnosis(necessarytoovercometheinterobservervariability)impossible.Basedonthestudiesonintergradervariabilityandevaluationofmachinelearningmodels,anadjudicatedimagediagnosiswasconsideredasthegroundtruth.[10]

Image analysis using AI-based offline softwareTheAI-basedautomated image analysis softwareused forthestudywasdesignedbyMediosTechnologies,Singapore,asubsidiaryofRemidioInnovativeSolutions.TheMediosAIalgorithmisbasedonconvolutionalneuralnetworks(CNN).TheAI consists of afirstneuralnetwork for imagequalityassessment and twootherneural networks thatdetectDRlesions.Thenetwork responsible for quality assessment isbasedon aMobileNet architecture. It consists of a binary

classifier andamessageprompts theuser to recapture theimageifitfailsthequalitycheck(QC).

Theneuralnetworkhasbeentrainedtoseparatehealthyfundus images (NoDR)with imageswith referableDR(definedasmoderateNPDRandabove).ThismaximizesthesensitivityforreferableDRandthespecificitytoruleoutallgradesofDR.Acomprehensivedatasetconsistingofimagestakeninavarietyofconditionshasbeenusedfortraining,withaproportionofittakenusingnonmydriaticand/orlow-costcameras. These include 4350 nonmydriatic images takenduringscreeningcampswiththeRemidioFundus-on-Phone;14,266imagescapturedwithaKOWAVx-10mydriaticcamera;and34,278 imagescomefromtheEyePACSdataset.Afinalper-patientDRdiagnosiswascomputedfromtheoutputsofthe neural networks and applied to all images of that patient. Apatientwasdeemedasreferableifthepredictionforoneormore images was positive.

The Medios software is integrated with the Remidio NM-FOP application loaded on the smartphone used toacquireimages.Thankstoleveragingonthehigh-performancecapabilities of the smartphonewithCoreMLandOpenGL,imageprocessingisdonedirectlyonthegraphicsprocessingunit(GPU)insteadofrelyingonaconnectiontoaserveronthe Internet.

In this study, theAI algorithmwas run offline by thetechnician on the smartphone itself after the imageswereacquired.The technicianwas trained to recapture images iftheAIgaveanalertof“poorimagequality.”TheAIQCwasfirst torunonimagesandthenthediagnosisof theAIwasrecordedasabinaryoutput,i.e.,DRpresentandNoDR.AllimagescapturedduringthisstudymetthequalitystandardsoftheAIandwereincludedintheanalysis.

Outcome measureThe primary aim of this study was to determine the sensitivity, specificity,positivepredictivevalue(PPV),andnegativePV(NPV)oftheAIalgorithmindetectingRDR(referableDRwasdefinedasmoderateNPDRormoreseveredisease,orthepresenceofDME)comparedwiththegoldstandarddiagnosisbyretinaspecialists.Thesecondaryaimsweretoassessthesensitivityandspecificityof theAIalgorithmin thediagnosisof“anyDR”andVTDR.AnyDRwasdefinedasmildNPDRormoreseverediseaseorthepresenceofDME,whileVTDRwasdefinedassevereNPDRormoreseverediseaseorthepresenceofDME.Anadjudicatedimagediagnosiswasconsideredasthegroundtruth.

Statistical analysisContinuousvariables arepresentedasmeanwith standarddeviation (SD) and categorical variables are presented asproportions(n,%).Thesensitivity,specificity,PPV,andNPVoftheAIinthedetectionofreferableDR,anyDRandVTDRwere calculated alongwith 95% confidence interval (CI).Theareaunder the receiveroperatingcurve (AUROC)wasplotted.Cohen’skappa(κ)wasmeasuredtoassessintergradervariability.All datawere stored inMicrosoft Excel andanalyzedusingtheStatasoftware(StataCorp14.2,Texas,USA).

ResultsThestudypopulationhadameanageof55±11years,durationofdiabetes11±8years,hemoglobinA1c(HbA1c)8±2%,and

[Downloaded free from http://www.ijo.in on Monday, January 20, 2020, IP: 106.206.50.176]

February2020 393Sosale, et al.: Offline smartphone‑based AI algorithm for DR diagnosis

bodymass index (BMI) 27 ± 4 kg/m2. Females constituted42% (n = 128) of the studypopulation. Thefinal analysisincludedimagesfrom297patients.Theimagesobtainedfromeitheroneorbotheyesofthe7(2.3%)patientswereconsideredungradable by the retina specialists and excluded. TherewasnoevidenceofDRin176participants(59.2%),butmildNPDRwasseenin35(11.7%),moderateNPDRin41(13.8%),severeNPDRin33(11.1%),andPDRin12(4%)patients.DMEwaspresentis36(20.4%)individualswithDRwithdifferentgradesofNPDRorPDR.Theintergraderagreementbetweentheretinaspecialistswas0.89forgradingofDRand0.9forgrading of DME.

The sensitivity and specificity of theAI algorithm indetecting RDRwas 98.84% (95% CI, 97.62–100%) and86.73% (95%CI, 82.87–90.59%), respectively,while thePPVwas 75.22% (95%CI, 70.31–80.13%) andNPVwas 99.46%(95%CI, 98.62–100%). TheAUROCwas 0.92 [Fig. 1].Anexample of the output from the Medios AI algorithm along withtherespectiveimagewithadiagnosisofreferableDRisshown in Fig.2.

For any DR, the sensitivity and specificity of theAI algorithmwas 86.78% (95% CI, 82.92–90.63%) and95.45% (95%CI, 93.09–97.82%), respectively, while thePPV andNPVwere 92.92% (95%CI, 90.00–95.84%) and91.30%(95%CI,88.10–94.51%).TheAUROCwas0.91[Fig.1].An example of the output from the Medios AI algorithm

Figure 1: Area under the receiver operating curve for referable diabetic retinopathy and any diabetic retinopathy

Figure 2: Example of the output from the Medios AI algorithm along with the respective image with a diagnosis of referable diabetic retinopathy

Figure 3: Example of the output from the Medios AI algorithm along with the respective image with a diagnosis of any diabetic retinopathy

Figure 4: Example of the output from the Medios AI algorithm along with the respective image with a diagnosis of proliferative diabetic retinopathy

alongwiththerespectiveimagewithadiagnosisofanyDRis shown in Fig.3.ThesensitivityforthediagnosisofVTDRwas100%[Fig.4].

Thenumberof falsepositiveswaseight (AIhad labeledeightcasesofNoDRashavingthedisease).Sixoutoftheseeight imageswere found tohaveartifacts thathadperhapsbeenmisidentifiedbythealgorithm.

DiscussionThis study aimed to evaluate theperformance ofMediosAI. It demonstrated that the AI algorithm has very high sensitivity and specificity todetectRDR, anyDR, aswellasVTDR comparedwithmanual, adjudicated grading byfellowship-trainedvitreoretinalsurgeons.Thealgorithmbeingintegratedwiththeimageacquisitionandstorageapplicationof an existing commercially available smartphone-basedimaging device, i.e., the RemidioNM FOP 10,made ituser-friendly. Seamless integrationof theAIwith theNMFOP10manufacturers’applicationmadetheimageworkflowsimpleandtime-efficientsothatthereportscouldbeproducedinreal-timebythetechnicianusingthedevice.

DRscreeningwithteleophthalmologyiswidelypracticedinIndiaandacrosstheworldwiththeuseofreadingcenters.Limitedaccesstotrainedreadersandinterobservervariabilityassociatedwithhumangrading led to thedevelopment ofautomatedsystemsforDR.Advancesindeeplearninghasled

[Downloaded free from http://www.ijo.in on Monday, January 20, 2020, IP: 106.206.50.176]

394 Indian Journal of Ophthalmology Volume68Issue2

to the development of several algorithms like the Google AI, EyeNuk,andIDx-DRforthedetectionofDR.TheCNN-basedsoftwaredesignedbyGooglewasusedonimagesfrompatientspresentingforDRscreeninginthreetertiaryeyehospitalsinIndia,anditshowedahighsensitivityandspecificity(>90%)fordetectingDR.[11]Inaprospectivestudy,thesensitivityandspecificityoftheGoogleAIforRDRatthefirststudysitewas88.9%(95%CI,85.8–91.5)and92.2%(95%CI,90.3–93.8);and92.1% (95%CI, 90.1–93.8) and95.2% (95%CI, 94.2–96.1) atsecondstudysite.[12]

Usinganotheralgorithmbasedondeepmachinelearningon thepublicly available fundus imagedatasets,Abramoffet al.alsofoundahighsensitivity(97%)andspecificity(87%),similar to our results.[13]InalargepivotalstudyconductedbyAbramoffet al.,thesensitivityandspecificityoftheIDx-DRsystem in identifying RDR met the United States Food and DrugAdministration(FDA)cut-offsforsuperiority.Thisstudymade IDx-DR thefirstFDAapprovedAIalgorithm for thediagnosis of RDR [Table1].[14]

Tufail et al. studied theperformance of threedifferentautomated image analysis software and reported a sensitivity andspecificity(>90%)usingtheEyeArt(EyenukInc.,WoodlandHills,CA)andRetmarker(Coimbra,Portugal)software.[15] The RetmarkersoftwarehasalsobeenusedtodetectDRinIndianeyes.[16] Roy et al.analyzed5780eyesof1445patientsthroughtheRetmarkersoftwareandfoundahighsensitivity(>90%)andlowspecificity(11–61%).Waltonet al.publishedoutcomeson the sensitivity and specificityof another algorithm, theIntelligentRetinalImagingSystem(IRIS)forDRscreening.Inthisretrospectivestudy,thesensitivityandspecificityofIRISwas66.4%and72.8%.[17]

TheEyeArt algorithmofEyenukhasbeen evaluated inseveral studies. Rajalakshmi et al. evaluated the EyeArt software forthedetectionofRDR[Table1]usingimagescapturedwiththe Remidio FOP.[18] Bhaskaranand et al.publishedresultsusingtheEyeArtsoftware,withmydriaticandnonmydriaticimagesfrom1,01,710eyes[Table1].[19]TheoutcomesfrommydriaticimagingweremarginallybetterwithimprovedsensitivityandgreaterAUROC.TheRemidioFOPhasnonmydriaticimageacquisitioncapabilitiesanditwillbeinterestingtoseehowtheAIalgorithmperformsonnonmydriaticimagesgoingforward.

Allthesoftwareprogramsthatarecloud-basedrequirehighcomputationalpowerandaboveall,Internetconnectivity,forreal-timereportingofresults.TheMediosAI,incontrastworksoffline,without Internet (or electricity).Toourknowledge,thisisoneofthefirstfewstudiesanalyzingtheaccuracyofanofflineAI-basedsoftwareforDRscreening.

ArecentstudybyNatarajanet al.,evaluatedtheperformanceof theMediosAIusing images captured from theRemidioFOP,in231individualswithdiabetesduringaDRscreeningprogram [Table1].Acommunityhealthcareworkercapturedtwo-segmentretinalimages.AnadditionalsensitivityanalysiswasperformedtoassesstheAI’sperformanceusingimagesthatarelikelytobecapturedduringlargescreeningcampsorhighpatientworkflowbytheunskilledworkforce.Bothgoodquality imagesand images thatdidnotmeet theminimumquality standardsby theAIwere included in this analysis.In the sensitivity analysis, the sensitivity of the AI for RDR remainedunchangedat100%,whilethespecificitydropped

from88.4% to 81.9%.Therewere a largernumber of falsepositivesoutputs (attributed topoor imagequality).Whilethismay lead to increased referrals,patient safetywasnotcompromisedastheAIdetectedallindividualswithRDR.ThisgivesanideaoftherealpracticaluseofAIforDRscreening.[20]

TheIDX-DRiscurrentlytheonlyFDAapprovedalgorithmforDRscreening.Whileseveralstudieshavebeenconductedtoevaluatevariousalgorithms,theonlyonescurrentlyinuseforDRscreeningcommerciallyintheUSareIDx-DR,andEyeArtintheEU;neitherareavailableinIndia.Whileweacknowledgethatthecomparisonofdifferentalgorithmsbasedonpublishedresults has its limitations (because ofdifferences in studymethods),wesummarizetheperformanceofthedeeplearningcloud-basedalgorithmscurrentlyusedforDRscreeningintheUSAandEU,andtheofflineMediosAI[Table1].

Intheirrecentpaperonthecurrentstateofteleophthalmologyin the United States, Rathi et al. describe applications ofteleophthalmologyinmanydiseases,includingDR.[7] Authors highlighttheupcomingroleofautomatedDRscreeningusingvariousalgorithmstoeasetheburdenofmanualDRscreening.AuthorsalsostatethatgiventheincreasingprevalenceofDR,theemergenceofautomatedscreeningservesasapromisingtooltoaddressthispublichealthissue.Inaddition,webelievethat it is extremely important tomakeupcomingAI-basedalgorithms offline forwidespread adoption.AnAI-basedalgorithm that gives consistent resultswithhigh accuracymayovercomehumanbarriers like inter-gradervariability,inadditiontoitsabilitytoprocessmillionsofimagesquickly,maybethebestwayforwardforgradingDRinthefuture.InIndia,andthedevelopingworldwithlimitedresources,whereaccesstotheInternetandcontinuouselectricityisachallengeinsmallertownsandvillages,thesetechnologiescanensurethatDRscreeningproceedsuninterrupted.

The advantages of this study are theuse of three-fieldphotography, grading, and adjudication by vitreoretinalsurgeonsasthegoldstandard.Thedrawbacksincludeasmallsample size and theuse of onlymydriatic images.Largerstudies,studieswithnonmydriaticimagingwilladdressiftheseresultscanbegeneralized.Studiesevaluatingtheintegrationof theAI into the clinicalworkflow, and comparisonwiththe clinicaldiagnosis froma comprehensive eye examandreal-worldstudieswillprovidemoreinsightandunderstandingofthistechnology.WeacknowledgethatthealgorithmisonlytrainedtodetectDRandcurrentlyworksonlyintegratedwiththeRemidioFOPcamera.AdditionalworktofocusongradingofDR,anddetectotherretinaldisordersisrequired.

Table 1: Performance of AI algorithms in the detection of referable diabetic retinopathy

AI software Sensitivity Specificity

FDA cut-off for superiority 85% 82.5%

Present Study Medios AI 98.8% 86.7%

Natarajan et al.[19] Medios AI 100% 88.4%

Rajalakshmi et al.[17] EyeArt 99.3% 68.8%

Bhaskaranand M et al.[18] EyeArt 91% 91%IDx[13] IDx-DR 87.2% 90.7%

AI=Artificial intelligence, FDA=United States Food and Drug Administration

[Downloaded free from http://www.ijo.in on Monday, January 20, 2020, IP: 106.206.50.176]

February2020 395Sosale, et al.: Offline smartphone‑based AI algorithm for DR diagnosis

ConclusionIn conclusion,ourpreliminary results show that thenovelAIalgorithmhashighsensitivityandspecificityindetectingRDRaswell asVTDR.This is probably the only softwareavailableinanofflinemodethatcandeliverresultsinstantlyin real-time. If used on a larger scale, it has thepotentialof ensuring timely referrals. The results of the SMART Study(SimpleMobile-BasedArtificialIntelligenceAlgorithminthediagnosisofDiabeticRetinopathy)withasamplesizeof900patientsusingnonmydriaticimagestotesttherobustnessof the Medios algorithm is awaited. Multiple larger studies thatshowreproducibilityandconsistencycanhelpvalidatethe algorithm further.We conclude that our findings areencouraging, furtherwork remains to improve the clinicalvalidity of these algorithms.

Acknowledgements and  supportWe acknowledgeMedios Technologies, Singapore, forprovidingtheAIsoftwareforconductingthestudyandFlorianM. Savoy,MediosTechnologies, for thedescriptionof thetechnicaldesignof theAIsoftware.Wealso thankRemidioInnovative SolutionsPvt. Ltd. forproviding theNMFOP10 camera to capture retinal images,Mr. Satish forhelpingwithdatacollectionandMrs.Roopa,thecameratechnician.Nofundingwasreceivedforthisstudy.

Financial support and sponsorshipNil.

Conflicts of interestMediosTechnologies onlyprovided theAI foruse for thestudy and played no role in the study design, funding, study process, data analysis, results or publication.Authors SB,SARare related to one of the two co-founders ofMediosTechnologies, Singapore.We have no personal financialinterests in the form of stocks and have received noremunerationorconsultationfeefromMediosTechnologies,Singapore.Theotherauthorshavenoconflictofinteresttodeclare.Nofundingwasreceivedforthisstudy.

References1. Gadkari SS,MaskatiQB,Nayak BK. Prevalence of diabetic

retinopathy in India: TheAll IndiaOphthalmological Societydiabeticretinopathyeyescreeningstudy2014.IndianJOphthalmol2016;64:38-44.

2. RamanR,RaniPK,ReddiRachepalleS,GnanamoorthyP,UthraS,KumaramanickavelG,et al.PrevalenceofdiabeticretinopathyinIndia:SankaraNethralayadiabeticretinopathyepidemiologyand molecular genetics study report 2. Ophthalmology2009;116:311-8.

3. Sengupta S, SindalMD, Baskaran P, PanU, Venkatesh R.Sensitivityandspecificityof smartphone-basedretinal imagingfordiabeticretinopathy:Acomparativestudy.OphthalmolRetina2019;3:146-53.

4. BastawrousA,RonoHK,Livingstone IA,WeissHA, JordanS,Kuper H, et al.Developmentandvalidationofasmartphone-basedvisual acuity test (PeekAcuity) for clinical practice andcommunity-basedfieldwork.JAMAOphthalmol2015;133:930-7.

5. RajalakshmiR,ArulmalarS,UshaM,PrathibaV,KareemuddinKS,Anjana RM, et al. Validation of smartphone based retinalphotography for diabetic retinopathy screening. PLoSOne2015;10:e0138285.

6. SharmaA,SubramaniamSD,RamachandranKI,LakshmikanthanC,Krishna S, Sundaramoorthy SK. Smartphone-based funduscameradevice(MIIRetCam)andtechniquewithabilitytoimageperipheralretina.EurJOphthalmol2016;26:142-4.

7. Rathi S, Tsui E,MehtaN,Zahid S, Schuman JS. The currentstate of teleophthalmology in the United States. Ophthalmology 2017;124:1729-34.

8. KapoorR,WaltersSP,Al-AswadLA.Thecurrentstateofartificialintelligenceinophthalmology.SurvOphthalmol2019;64:233-40.

9. WilkinsonCP,FerrisFL,KleinRE,LeePP,AgardhCD,DavisM,et al. Proposed international clinical diabetic retinopathy anddiabeticmacularedemadiseaseseverityscales.Ophthalmology2003;110:1677-82.

10. KrauseJ,GulshanV,RahimyE,KarthP,WidnerK,CorradoG,et al.Gradervariabilityandtheimportanceofreferencestandardsforevaluatingmachinelearningmodelsfordiabeticretinopathy.Ophthalmology2018;125:1264-72.

11. GulshanV,PengL,CoramM,StumpeMC,WuD,NarayanaswamyA,et al. Development and validation of a deep learning algorithm for detectionofdiabeticretinopathyinretinalfundusphotographs.JAMA2016;316:2402-10.

12. GulshanV,RajanR,WidnerK,WuD,WubbelsP,RhodesT, et al.Performanceofadeep-learningalgorithmvsmanualgradingfordetectingdiabetic retinopathy in India. JAMAOphthalmol2019;137:987.

13. AbràmoffMD,LouY,ErginayA,ClaridaW,AmelonR,FolkJC,et al.Improvedautomateddetectionofdiabeticretinopathyonapubliclyavailabledataset throughintegrationofdeeplearning.InvestOphthalmolVisSci2016;57:5200-6.

14. AbramoffMD,LavinPT,BirchM,ShahN,FolkJC.Pivotaltrialof an autonomousAI-based diagnostic system for detectionofdiabetic retinopathy inprimary care offices. npjDigitMed2018;1:39.

15. TufailA,KapetanakisVV, Salas-Vega S, EganC,RudisillC,OwenCG, et al.Anobservational study toassess if automateddiabeticretinopathyimageassessmentsoftwarecanreplaceoneor more steps of manual imaging grading and to determine their cost-effectiveness.HealthTechnolAssess2016;20:1-72.

16. RoyR, LoboA, LobA, Pal BP,OliveiraCM,RamanR, et al. AutomateddiabeticretinopathyimaginginIndianeyes:Apilotstudy.IndianJOphthalmol2014;62:1121-4.

17. WaltonOB,GaroonRB,WengCY,GrossJ,YoungAK,CameroKA,et al.Evaluationofautomatedteleretinalscreeningprogramfordiabeticretinopathy.JAMAOphthalmol2016;134:204-9.

18. RajalakshmiR,SubashiniR,AnjanaRM,MohanV.Automateddiabetic retinopathy detection in smartphone-based fundusphotographyusingartificialintelligence.Eye2018;32:1138-44.

19. BhaskaranandM,RamachandraC,BhatS,CuadrosJ,NittalaM,Sadda S, et al. The value of automateddiabetic retinopathyscreeningwiththeEyeArtsystem:Astudyofmorethan100,000consecutive encounters frompeoplewith diabetes.DiabetesTechnolTher2019;21:635-43.

20. NatarajanS,JainA,KrishnanR,RogyeA,SivaprasadS.Diagnosticaccuracyofcommunity-baseddiabeticretinopathyscreeningwithanofflineartificial intelligencesystemonasmartphone. JAMAOphthalmol2019.doi:10.1001/JAMAOphthalmol.2019.2923.

[Downloaded free from http://www.ijo.in on Monday, January 20, 2020, IP: 106.206.50.176]

top related