Neuropathy Classification of Corneal Nerve Images Using Artificial Intelligence
Advisor | AlMaadeed, Sumaya |
Author | Salahuddin, Tooba |
Available date | 2019-12-09T10:14:04Z |
Publication Date | 2019-06 |
Abstract | Nerve variations in the human cornea have been associated with alterations in the neuropathy state of a patient suffering from chronic diseases. For some diseases, such as diabetes, detection of neuropathy prior to visible symptoms is important, whereas for others, such as multiple sclerosis, early prediction of disease worsening is crucial. As current methods fail to provide early diagnosis of neuropathy, in vivo corneal confocal microscopy enables very early insight into the nerve damage by illuminating and magnifying the human cornea. This non-invasive method captures a sequence of images from the corneal sub-basal nerve plexus. Current practices of manual nerve tracing and classification impede the advancement of medical research in this domain. Since corneal nerve analysis for neuropathy is in its initial stages, there is a dire need for process automation. To address this limitation, we seek to automate the two stages of this process: nerve segmentation and neuropathy classification of images. For nerve segmentation, we compare the performance of two existing solutions on multiple datasets to select the appropriate method and proceed to the classification stage. Consequently, we approach neuropathy classification of the images through artificial intelligence using Adaptive Neuro-Fuzzy Inference System, Support Vector Machines, Naïve Bayes and k-nearest neighbors. We further compare the performance of machine learning classifiers with deep learning. We ascertained that nerve segmentation using convolutional neural networks provided a significant improvement in sensitivity and false negative rate by at least 5% over the state-of-the-art software. For classification, ANFIS yielded the best classification accuracy of 93.7% compared to other classifiers. Furthermore, for this problem, machine learning approaches performed better in terms of classification accuracy than deep learning. |
Language | en |
Subject | chronic diseases early prediction - chronic diseases |
Type | Master Thesis |
Department | Computing |
Files in this item
This item appears in the following Collection(s)
-
Computer Science & Engineering [100 items ]