Date of Submission
Spring 5-9-2017
Degree Type
Thesis
Degree Name
Master of Science in Computer Science (MSCS)
Department
Computer Science
Committee Chair/First Advisor
Dr. Jose Garrido
Track
Others
Convolutional Neural Networks
Chair
Dr. Dan Chia-Tien Lo
Committee Member
Dr. Jose Garrido
Committee Member
Dr. Dan Chia-Tien Lo
Committee Member
Sumit Chakravarty
Abstract
Convolutional Neural Networks (CNNs), the state of the art in image classification, have proven to be as effective as an ophthalmologist, when detecting Referable Diabetic Retinopathy (RDR). Having a size of less than 1\% of the total image, microaneurysms are early lesions in DR that are difficult to classify. The purpose of this thesis is to improve the accuracy of detection of microaneurysms using a model that includes two CNNs with different input image sizes, 60x60 and 420x420 pixels. These models were trained using the Kaggle and Messidor datasets and tested independently against the Kaggle dataset, showing a sensitivity of 95\% and 91\%, a specificity of 98\% and 93\%, and an area under the Receiver Operating Characteristics curve of 0.98 and 0.96, respectively, in the sliced images. Furthermore, by combining these trained models, there was a reduction of false positives for complete images by about 50\% and a sensitivity of 96\% when tested against the DIARETDB1 dataset . In addition, a powerful image pre-processing procedure was implemented, which included adjusting luminescence and color reduction, improving not only images for annotations, but also decreasing the number of epochs during training. Finally, a novel feedback operation that re-sent batches not classified as well as expected, increased the accuracy of the CNN 420 x 420 pixel input model.