Collaborative Data Collection for Bio-Imaging Algorithms
  • No tags were found...

Collaborative Data Collection for Bio-Imaging Algorithms

Collaborative Data Collection forBio-Imaging AlgorithmsGabe CastanonMentor: John O’DonovanFaculty Advisor: B.S. ManjunathCenter for Bioimage Informatics

Goals:To create a “social web” application inwhich users provide information.Social WebUserWebMasterOld WebWebsiteWebsiteUser

GoalsThe key idea behind the application is that“community wisdom” can be applied tosolve problems within the ability ofcommunity members.Specifically, developers of imageprocessing algorithms require“ground truth” data to test the algorithms.

Goals:Ground truth is a point of interest which ismanually found. It is used as a control totest Image processing algorithms against.Gathering this data is an easy but timeconsumingand tedious task.

Goals:Proposed Solution:Accurate and reliable ground truth data canbe acquired through a collaborative system.More specifically, an application to allowusers on Facebook to provide “groundtruth” on biological images in the form ofgraphical annotations (user created points).

ApplicationThe social web application Facebook was chosenbecause of its large user base (80,000,000).Facebook also provides an API to allow for userprovided applications.

ApplicationSharing /PersonalizationAdministratorControlApplicationPageAnnotator/ViewerProfilePageWidgetControlPanel

Application ArchitectureDatabase(MySQL)Php sql driverApache Web Server (mod-php)Clusteringengine(DBSCAN)Data Server(PHP)AppilcationController(php/js)AJAX (JSON)AJAX (JSON)Facebook Application ServerAdministratorInterface(DHTML/FBML/EmbeddediFrames)ApplicationInterface(DHTML/FBML/EmbeddediFrames)FBMLFacebookDatabaseFacebookAPI

Database SchemaTABLE:annotation_dataEntryID | int(11)x | floaty | floatusername | varchar(60)imageid | int(11)date | varchar(50)time_taken | int(11)TABLE:feature_datafeatureID | int(11)x | float(11)y | float(11)imageID | int(11)TABLE:user_datauserid | int(11)firstname | varchar(50)lastname | varchar(50)city | varchar(50)state | varchar(50)zip | int(11)country | varchar(50)ed_name | varchar(50)ed_conc | varchar(50)hs_name | varchar(50)

Annotation/Image ViewerWithin the annotator, users click points onthe randomly generated image. Theapplication then stores the point in the formof X and Y coordinates.

Annotation/Image ViewerWith the collected annotations, there areoptions to visualize the data in various ways.

DataOnce the application has been open tothe public for a longer period of time,we hope to see data clusters aroundareas of interest.clustered data on the annotator

Data ClusteringThe GroundTruth application gathersinformation from many users on Facebook.To use collaborative information effectively, itmust be combined in some way.In our application this is done with aclustering algorithm which is runthrough the administration console.

Data ClusteringMy algorithm is loosely based on the DBSCANalgorithm (Density-Based Spatial Clustering ofApplications with Noise). (Ester et al. 1996)

Data ClusteringThe algorithmthen averages thecoordinates of thepoints within eachcluster.MinPts = Theminimum number ofpoints a candidate setmust have to beconsidered a cluster.MinDistThreshold

ResultsAfter sufficient annotations are gathered, thedata can be used to test automated imageprocessing algorithms.Recap: By averaging over a large amount ofcollaborative input, accurate ground truth datais obtained.

Results•Ground truth has been established on a set of retinalimages from rats. The ends of dendrites in 10 retinaldetachment images have been correctly identified.•Currently: Approximately 400 collaborative annotationsfrom 7 users (Expect an order of magnitude more whenFacebook registers the application).•Average time spent per annotation is: 8430 millisec.•From this data, DBScan clustering algorithm has found66 reliable clusters.•The center points of each of these clusters isconsidered as a feature.

FutureThe application is extensible and can be used tocollect manual ground truth information for abroad range of automated tasks.

In RetrospectThroughout this program, I have spent time withgrad-students and post-docs. I learned what it islike to continue in academia after obtaining anundergraduate degree, and will use this to myadvantage when deciding what I want my future tobe.Not only have I been immersed in a universitysetting, I have also gained the knowledge andexperience necessary to continue my academiccareer.

Acknowledgments• I would like to take this opportunity to thankeveryone involved in the AR program, including:• John O’Donovan• Lubi Lenaburg• Evelyn , Anthony, and Leila• The Allison family• Prof. B.S. Manjunath

More magazines by this user
Similar magazines