Please use this identifier to cite or link to this item:
Title: Detecció automàtica de capes retinals en imatges OCT
Author: Rosés Castellsaguer, Jordi
Director: Ventura Royo, Carles  
Tutor: Nuñez Do Rio, Joan Manuel
Keywords: OCT images
image segmentation
convolutional neural networks
Issue Date: 12-Jun-2019
Publisher: Universitat Oberta de Catalunya (UOC)
Abstract: Optical Coherence Tomography (OCT) is a non-invasive technology that allows images of ocular tissues to be obtained. Given its usefulness, it is necessary to develop diagnostic support tools that help automatically process the image in order to facilitate its analysis and to make retina features more evident. In this context, our proposal is to make a comparative study between three automatic processes for detection of retinal layers in OCT images, all of them transforming an image segmentation problem into a pixel classification one: 1) Border detection with two classes ("no border" vs "border") and using local first and second order statistics as entries to a Random Forest type classification model. 2) Border detection, also with two classes, and using sub-images as inputs to a Convolutional Neural Network (CNN) model. 3) Classification of the pixels in the image between multiple regions ("image background", "inner eyeball", "boundaries between layers", "retinal layer 1", "retinal layer 2", etc.), and using sub-images such as inputs to a Convolutional Neural Network (CNN) model. We will see that better results are obtained by considering the exercise as a two-class problem (with AUC values above 0.95). Even though the multi-class option does not yield as good classification results, it eliminates the necessity of any subsequent result processing.
Language: Catalan
Appears in Collections:Bachelor thesis, research projects, etc.

Files in This Item:
File Description SizeFormat 
jrosescTFM0719memòria.pdfMemòria del TFM2.43 MBAdobe PDFView/Open
jrosescTFM0719presentació.pdfPresentació del TFM1.17 MBAdobe PDFView/Open
UOC_TFG_Jordi_Roses_VIDEO.m4v133.59 MBMP4View/Open

This item is licensed under a Creative Commons License Creative Commons