PortableRetinaEyeScanningDevice
1
EngineeringScienceDepartmentSonomaStateUniversityStudents:CristinFaria&DiegoA.EspinosaFacultyAdvisor:Dr.SudhirShresthaIndustryAdvisor:BenValvodinosClient:NorthBayVisionCenter
Website:http://diabeticretinopathyssu.weebly.comEmail:[email protected]@sonoma.edu
1. ProblemandSolution2. GeneralSystemOverview3. SystemDescriptionandTechnicalComponents4. MarketingandEngineeringRequirements5. TestingandResults6. MaterialsandCosts7. Timeline8. Challenges9. LessonsLearned10. FutureWork11. Conclusion
2
Overview
3
Problem
• Millionsofpeopleintheworldarediagnosedwithdiabeteseveryyear.
• Leadingtodiabeticretinopathy,adiseasefoundintheretinaoftheindividual.
• Oftenleadingtoblindnessifleftuntreated.
• Thisdiseaseisnoteasilydetected,andnormallynotreversiblewhenfoundinthepatient.
4
PreviousWorksandourSolution
PeekVision:● Createdcamerasandsoftwareapplicationsonsmartphonestoscanpatients’retinas,butdoes
notutilizemachinelearningsorunsslowly.
Epipole:● Handheldretinalfunduscamera,mustbeconnectedtoWindowsorAndroidandtheInternet.
OurSolution:● AstandaloneproductthatisabletogiveinstantresultsbyutilizingMachineLearningand
ImageProcessing.● NoInternetconnectionnecessary.
Ourdeviceisadiagnostictool,andresultsgivenfromdeviceshouldbetakentoalicensedmedicalprofessional
5
FinalProduct
● Imageofourfinalproduct.
● Howcomputer,raspberrypi,camera,LEDandbatteryareallconnected.
Explainalltheparts– diagram
6
SystemOverview
● Systemoverviewthatlooksoverentiredeviceandshowshoweachpieceisconnected.
● Basicoverviewofhowoursystemworksandcanbeimplemented.
7
HardwareDesign
● Visualrepresentationofallofthehardwarecomponentsweutilized.
● RaspberryPiisthecentralunit.
● Demonstrateshowallcomponentsinteractwitheachother.
8
SoftwareDesign
● Thisflowchartgivesageneraldescriptionofwhatourprogramdoes.● Initializes,takesapicture,determineswhetherornottheretinaishealthyand
displaystheresultviaLED.● ThisprogramutilizesthemachinelearningtoolboxinMATLAB,RaspberryPi3
ModelB,andLEDtogiveaccurateresults.
Requirements
MarketingRequirements● Thedevicewillbereliableindeterminingthe
healthstatusofapictureofapatient'sretina.● Thedevicewillnotbefullyautomated,butwill
besimpletouse.● Thedevicewillcapturepicturesofaretina
fundusimageswithhighresolution.● Thedevicewillbemarketedtowardtrained
physiciansandhealthpractitionersinvolvedintreatingpatientsinruralregionswithlimitedaccesstoophthalmologytestingtechnologythroughouttheworld.
9
EngineeringRequirements● Theaccuracyofthescanningmustbeover90%
inorderforthedevicetobereliableindeterminingthestatusofthepicture.
● Thedevicewillconsistoffourcomponents:RaspberryPi3ModelB,ArduCAM(camera),BatteryPack,andLEDboxlight.
● TheprojectrequiresinterfacingMATLABwiththedevicebut,minimalcommandsareusedtoexecutethedevice.
● Functionalityofthedevicewillbeaprecursortotakingascanofapatient’sretina,apracticedonebylicensedindividuals.
Foracompletelistofourrequirementsvisitourwebsite:http://diabeticretinopathyssu.weebly.com
10
MachineLearningandtheConfusionMatrix
WhatisMachineLearning?● Itisatypeofcomputerprogrammingwhichusesdatatoperformatask.● Themoredata,orimagesyouaddtoaprogram,thebetteritperforms.
WhatistheConfusionMatrix?● Itisatablethatisusedtodescribetheperformanceofaclassifier(methodusedtoclassify
data)onasetoftestdatawherethetruevaluesareknown.● Inourprojectthemostimportantpartofthematrixlookedatthefalsenegativerateand
truepositiveratetodeterminetheoverallaccuracyofouralgorithm.
11
MachineLearningandourProject
● Ourprojectutilizesmachinelearningandimageprocessinginordertorunthealgorithm,anddifferentiatebetweenimages.
● Inordertocheckbetweenhealthyretinaimagesanddiabeticretinaimagesouralgorithmlooksatbloodvessels,darkspots,damagetotheretinashapetounderstandthedifferencesbetweenthetwocategories.
12
ClassifierTest
● Testusedtodeterminewhichtypeofclassifierwouldbebestsuitedtotestouralgorithmmostaccurately.
● Tested21differentclassifiersandfoundthat8hadsamebestaccuracyof60%.
● Thistestwascompletedbeforeweimprovedtheaverageaccuracyofouralgorithm.
● LookedatconfusionmatricesandconcludedthatLinearSupportVectorMachine(SVM)wouldgiveusthebestresult.
13
AccuracyImprovementTest
● InitiallythebestaccuracyfortheLinearSupportVectorModelwas60%whichwas30%lowerthanourminimumgoalaccuracyof90%.
● ToincreaseouraccuracyfirstweeliminatedourGlaucomagroupfromourcode,thisincreasedouraverageaccuracyto83%.
● IncreasedthenumberofimagesinDiabeticRetinopathyfrom15to45,thisactionincreasedouraverageaccuracyto91.7%.
● OurFalsediscoveryrateforDiabeticRetinopathyis2%andis22%forhealthyretinas.
14
AlgorithmAccuracyTest
● Testedtheaccuracyofthealgorithm,bytesting60images,30withdiabeticretinopathyand30healthyretinalimages.
● Imagestestedwerenotincludedinourtestingalgorithm,weretakenbyourcamera,andhadnotbeenpreviouslyseenbyouralgorithm.
● Wewereabletoconclude5healthyimagesweregivenaninaccurateresult,andonlyasingleimagewithdiabeticretinopathywasgivenaninaccurateresult.
● Givenanaverageaccuracyof90%,diabeticretinopathyaccuracyof97%,andhealthyaccuracyof83%.
● Diabeticretinopathyhadafalsediscoveryrateof3%andhealthyhadafalsediscoveryrateof17%.
15
AutomationTest
● Automatedtakingpictures,savingthem,andtestingthem.
● Itautomaticallytakeimageswithanexternalcamera,savesthemtoaspecificlocation,andteststhemwithouralgorithm.
● Thisprogramworkswithasingleclickofamouse.
16
HealthIndicatorTest
● TestedtoconfirmthattheLEDwascompatiblewithMATLABandwouldgiveusthecorrectresult.
● IftheLEDturnsonitmeanstheimagehasdiabeticretinopathy,ifnottheimageishealthy.
● TestedImageresultisDiabeticRetinopathy.
17
MaterialsandCosts
● SpecialThankstoSOURCEwhogaveus$400tocompleteourproject.
● SpecialthankstoSharahmMarivaniforprovidingcomponents.
18
Timeline
● TestingandDebuggingourprojecttooklongerduetocameraandRaspberryPiissues.
● Communicationledtosomedrawbacksandtimeconstraints.
● GettingMATLABtocommunicatecorrectlywithRaspberryPicauseddelays.
● Evenwiththeseissueswewerestillabletomakeacompletedproject.
• WehadtochangetheideaofourprojectfromNVIDIAtoRaspberryPiduetolackofNVIDIAsupport.
• Utilizingmachinelearningandimageprocessing.
• GettingMATLABonRaspberryPi.
• Testingandupdatingouralgorithmtogetittoover90%accuracy.
19
Challenges
● Progressisnotastraightpath,especiallywhenyouhaveaprojectthatismainlysoftwarebased.
● Thingsgowrong,andbythingswemeaneverything.
● Teamwork,anddedicationarevitalwhentakingonanewproject.
20
LessonsLearned
21
FutureWork
● The next phase of our project will be to attach a retinal fundus camera to our device in order to take an image of a patient’s retina.
● Introducing, the product to doctors and health practitioners in regions that to do not have access to this technology, and make it so patients in these different regions have access to better medical care and are able to get treated faster.
22
SpecialThanks
● We would like to thank our Professor Dr. Farid Farahamand, Faculty Advisor Dr. Sudhir Shrestha, Industry Advisor Ben Valvodinos, and Client North Bay Vision Center for giving us feedback, and helping us with the completion of our project.
● We would also like to thank our family and friends, for their encouragement and support throughout our senior design project, those actions and kind words did not go unnoticed.
• https://www.aoa.org/patients-and-public/eye-and-vision-problems/glossary-of-eye-and-vision-conditions/diabetic-retinopathy
• http://www.diabetes.co.uk/news/2014/may/portable-eye-scanner-to-revolutionise-detection-of-diabetic-retinopathy-96133928.html
• http://www.who.int/mediacentre/news/releases/2003/pr86/en/• https://www.mathworks.com/discovery/machine-learning.html• https://www.sas.com/en_us/insights/analytics/machine-learning.html#• https://www.engineersgarage.com/articles/image-processing-tutorial-
applications• https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4463765/• http://onlinelibrary.wiley.com/doi/10.1046/j.1464-5491.2000.00333.x/full
23
References
Questions/Comments
24
Questions
25
ClassifierTestTable1
● Took 21 classifiers and looked at the average accuracy using the confusion matrix.
● Found that 8 different classifiers had same initial accuracy of 60%.
● These accuracies were our initial accuracies before adding anything to our algorithm.
26
ClassifierTestTable2
● Looked at the True Positive Rate and False Negative Rate of top 8 classifiers.
● Was able to conclude from the test that Linear SVM gave us the best overall accuracy rating in each category.
● False Negative Rate was lower in all three categories and gave us a more well-rounded accuracy.
27
AlgorithmImprovementTestMatrix1
● Single confusion matrix for Linear Support Vector Model.
● The top row specifies Diabetic Retinopathy, where 41 out of 45 images were correctly categorized, with an accuracy rating of 91%.
● Bottom row specifies the healthy category where 14 out of 15 were categorized correctly with an accuracy rating of 93%.
● Which gave an average accuracy of 92%.
28
AlgorithmImprovementTestMatrix2
● Confusion matrix for Linear Support Vector Model, that shows true positive and false negative rates.
● For Diabetic retinopathy the true positive rate was 91% and false negative rate was 9%.
● For Healthy the true positive rate was 93% and false negative rate was 7%.
29
AlgorithmAccuracyTestGraph
● BluebarisDiabeticRetinopathyandorangeisHealthy.
● Tested30imageswithDiabeticRetinopathy,andwasgiven29correctresults.
● Resultedina3%Falsenegativerate.
● Tested30HealthyImagesandwasgiven25correctresults.
● Resultedina17%Falsenegativerate.
30
AutomationTest
● Codeusedinordertoautomateourprogram.
● WritteninMATLABandutilizesourmachinelearningalgorithm.