Purchase this article with an account.
A.B. Williams, G. Thomas, M.A. Grassi, J.R. Lee; IDOCS: Collaboratively Describing and Classifying the Features of AMD with Machine Learning . Invest. Ophthalmol. Vis. Sci. 2006;47(13):2096.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Predisposition to Age–related Macular Degeneration (AMD) is associated with a number of specific genes. A major obstacle in categorizing these genes is an inability to reliably stratify patients with different molecular subtypes into homogenous groups. Our goal is to use the power of multi–agent systems computer technology to facilitate the collaborative development of a robust AMD classification system using the Intelligent Distributed Ontology Consensus System (IDOCS).
100 fundus images of patients with AMD were obtained from the Ophthalmology clinic at the University of Iowa. 3 geographically distributed clinicians with AMD expertise described the features in digital voice recorders. We manually generated an AMD vocabulary by analyzing the data to extract a superset of all AMD features and their corresponding attributes. These AMD features and attributes were incorporated into IDOCS. IDOCS provided a web–based user interface that allowed for the description of AMD features in conjunction with a capability to annotate corresponding areas of the digital fundus image. A second group of 100 patients was selected, each with multiple supporting images. Using the IDOCS interface, the clinical experts described the AMD features and attributes contained in the images. Machine Learning with 10–fold cross validation was performed on the data of each expert using Decision Trees (C4.8), Bayes Net Classifiers, Locally Weighted Learning and Bagging. The data of each clinician was analyzed using the classifiers of the other two; evaluation of the generated confusion matrices was performed. To validate the observations, we clustered the combined data using the EM Clustering algorithm.
The percentage of images correctly classified varied from 70% to 82%. The analysis suggests that there are 4 primary mappings among drusen described by the experts. The exact probability of similarity and the degree of overlap between feature definitions of our clinicians vary based on the learning classifier used, but each collaborator’s drusen features have 2–4 clear counterparts with that of the other clinicians.
In summary, we created and tested the use of machine learning in AMD classification through the use of IDOCS. Future work includes using the precise image data to improve matches between collaborators and using fuzzy logic to fine tune the similarity and boundary determination.
This PDF is available to Subscribers Only