Artikel
Developing human based intuitive deep learning algorithms for analyzing intermediate AMD OCT-images
Suche in Medline nach
Autoren
Veröffentlicht: | 5. Februar 2020 |
---|
Gliederung
Text
Background: While automated evaluation procedures exist to quantitate various anatomic aspects of optical coherence tomography (OCT), OCT analysis of eyes with intermediate age-related macular degeneration (iAMD) often requires verification or modification by a trained OCT evaluator. Newer modalities including various machine learning approaches are under investigation in an attempt to automate the analysis as well as understanding algorithmic approaches to defining anatomic characteristics that correlate with iAMD progression.
Methods: Human computing has proven to be an effective way to crowdsource a variety of scientific problems, as well as to leverage human pattern-recognition ability. Video games allow users to interact with the scientific data while also leveraging the elements game developers require to maintain engagement. To investigate whether game interactions can train players to evaluate iAMD OCT images, a web-based game, Eye in the Sky: Defender, was created featuring gameplay designed around quantification of drusen.
Results: Evaluations of accuracy using the mean user line input reflected 86% improvement from a players' initial image evaluation. Spearman rank correlation and Procrustes analysis indicate mean line accuracy within 10% margin of error by image 4 and improved Results compared to the automatically generated line in more challenging images. The preliminary result of this approach allowed the development of a human intuition filter that, when combined with standard machine learning, improved precision and accuracy of drusen identification with fewer OCT image inputs.
Conclusions: These Results suggest human computation games can be used to expedite algorithmic development of iAMD OCT analysis.