Year: 2024
Type: dissertation
Abstract: In this thesis, different ways of segmenting retinal layers in OCT images are investigated. This is done through experiments, where a U-net, a lightweight fully convolutional network and two approaches based on the Canny edge detector are implemented and evaluated using Dice score, precision, recall, and Hausdorff distance. Bioptigen, Heidelberg, and Zeiss are examined as vendors and their metrics are compared separately in addition to the metrics for all data. Additionally, the performance was evaluated separately with images without denoising, images denoised using Gaussian smoothing, and non-local means respectively. The convergence time for the neural networks is also examined. The results show that the neural network approaches perform well on all vendors and denoising approaches, while the Canny edge detectors achieve comparatively worse results, showing less suitability to solving the segmentation problem. The U-net architecture achieved the best performance with regards to the Dice score, precision, and recall with a mean scores of approximately of 0.99. On the other hand, the lightweight FCN achieved the best values for Hausdorff distance, with a mean Hausdorff distance of 4.943 when using Gaussian smoothing. Among the vendors, the approaches performed best on the Zeiss data, then the Bioptigen data, and lastly the Heidelberg data. The reason for this could be that the Zeiss images that were used contained less noise due to them being preprocessed, and the Heidelberg data sets contained most of the difficult cases with for example DME. Out of the three denoising approaches, Gaussian smoothing achieved the best results but non-local means resulted in the fastest convergence time for the neural networks.
Source: DiVA (Linkoping University)
Author: Ohlsson Orell, Julia
Institution: Linköping University
Language: English
Topic: Retinal Imaging and Analysis
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: green
Funder: Linköpings Universitet
OpenAlex ID: W7025158384
