QUANTIFYING LOW-LEVEL IMAGE FEATURES FOR RETINAL IMAGE SEGMENTATION

UoM administered thesis: Phd

  • Authors:
  • Qinhao Wu

Abstract

Inspired by the morphological study of conditions that can cause blindness: glaucoma and diabetic retinopathy, this research project aims to offer more precise pattern segmentation on retinal images for assisting diagnosis. Specifically, colour and texture features were proposed to improve the pattern segmentation. The research project started with the analysis of intensity changes in the retinal image, leading to the proposal of a method for optic nerve head segmentation using the colour features. Due to the unbalanced intensity and sensitivity to noise, the texture feature was used to handle local noise and offer precise segmentation results, by the Binary Robust Independent Elementary Features (BRIEF). Moreover, BRIEF was enhanced by extending it to all colour channels, called CBrief, resulting in a texture descriptor whose performance is comparable with the state-of-the-art. In testing the performance of segmentation, CBrief achieved Accuracy = 93.4%, Sensitivity = 72.6%, and Specificity = 95.1% in the texture synthesised vascular test. However, CBrief failed to extract the colour-texture feature from retinal images. In order to investigate the texture in retinal images, the deep texture descriptor, FVCNN, was applied. The result showed that deep texture descriptor could help in distinguishing the optic nerve head, blood vessels, and background. To draw the conclusion, with the study of colour and texture, the new colour-texture descriptor CBrief was proposed and achieved outstanding performance in texture classification and segmentation. However, due to the subtlety of the texture contained in retinal images, it is hard to extract useful texture information. However, the result of FV-CNN suggested the potential of using deep texture information on the deep segmentation model.

Details

Original languageEnglish
Awarding Institution
Supervisors/Advisors
  • Tim Morris (Supervisor)
  • Carole Twining (Supervisor)
Award date1 Aug 2020