ETD system

Electronic theses and dissertations repository

 

Tesi etd-11202019-005720


Thesis type
Tesi di laurea magistrale
Author
FAGNANI, ALESSANDRO
URN
etd-11202019-005720
Title
Deep learning-based segmentation of retinal layers in intraoperative Optical Coherence Tomography
Struttura
INFORMATICA
Corso di studi
INFORMATICA
Commissione
relatore Bacciu, Davide
Parole chiave
  • segmentation
  • deep learning
  • segmentazione
  • retinal layer
  • optical coherence tomography
  • machine learning
Data inizio appello
06/12/2019;
Consultabilità
secretata d'ufficio
Data di rilascio
06/12/2022
Riassunto analitico
This thesis studies the approaches for the segmentation of retinal layers in Optical Coherence Tomography (OCT) scans. It proposes an overview of the technology and the medical applications of these systems, with a focus on the advantages of intraoperative systems and, in particular, the added progress of having an automatic retinal layers' segmentation algorithm. It has been proved in these last years, that Deep Learning can deal efficiently computer vision tasks and there are examples in literature on using these recent methods to segment retinal layer on diagnostic OCT scans. In this thesis, it is presented a recap of the most recent and successful between these approaches and the technical ideas behind them. However, to the knowledge of the author, there are no segmentation algorithms applied to intraoperative systems. This study presents a model, specifically used for edge detection, in order to find two retinal layers, Inner Limiting Layer (ILM) and Retinal Pigment Epithelium (RPE), in intraoperative OCT B-scans. The task of segmenting retinal layers in intraoperative OCT scans is more challenging due to the noisy and (sometimes) unreliable nature of the images. Moreover, it is also presented an extension of the model with an adversarial extension, to avoid specific issues of the task and improve the outcome. Along with the results, the study shows a qualitative analysis for a more complete evaluation.
File