Image analysis on hyperspectral corneal epithelium injuries using a new image enhancement and a novel mixing kernel

Student thesis: Doctoral Thesis

Abstract

The work presented in this thesis is to improve the evaluation of corneal injuries with the aim of eliminating the application of eye stains. In the case of this study, a new algorithm of image enhancement, mixture deep learning classification, and a novel mixing kernel for support vector machine are proposed to provide objective measurements. The proposed algorithms of image enhancement and the mixing kernel in this study are to provide a new contribution to knowledge in the field of ophthalmology and machine learning. To the authors’ knowledge, this is the first analysis of 25 corneal epithelium of porcine eyes using hyperspectral imaging combined with image processing algorithm. In addition, four series of experiments on the cornea were carried out with and without eye staining through the HSI scanning method. The analysis mainly focused on the eyes without staining, while the stained eyes were used as the ground truth images for the purpose of identifying the injured area.In t his study, a new 8-Step Hyperspectral Image Enhancement (8-SHIE) was developed to differentiate the injured and healthy corneas. The results showed that the proposed algorithm was able to clearly highlight the boundary of the injury. This algorithm was further tested on the existing remote-sensing Indian Pines dataset in order to ascertain that it can work well with other hyperspectral images. Moreover, the algorithm can also be used to monitor the cornea healing process considering that the injured boundary can be viewed from band-to-band. Overall, it can be concluded that this algorithm is able to successfully separates the ten varying classes despite its main purpose of distinguishing the two classes. All the enhanced images are then classified using mixture deep-learning technique.This study also introduced a novel design of mixing kernels (3-ConvSvm) for binary classification in support vector machine classifier. Three standard kernels, namely RBF, polynomial, and RQKcombined in order to provide more learning flexibility due to the various parameters setting. The algorithms proposed aim to minimise the generalisation error during the classification when one or more parameters is tuned. Apart from that, the numerical experiments also showed that the new kernel performs similarly and sometimes even better than the standard kernels (RBF, polynomial, RQK, mixed two-kernels). Finally, the results revealed that the performance of 3-ConvSvm able to reduce the error loss by tuned the kernel parameters.
Date of Award1 Mar 2018
LanguageEnglish
Awarding Institution
  • University Of Strathclyde
SupervisorStephen Marshall (Supervisor) & Jinchang Ren (Supervisor)

Cite this