Donders Institute for Brain, Cognition and Behaviour
Zoek in de site...

Thesis defense Bart Liefers (Donders series 575)

21 September 2022

Promotors: prof. dr. ir. C. Sánchez Gutiérrez, prof. dr. B. van Ginneken, and prof. dr. C. Hoyng
Copromotor: dr. T. Theelen

Deep learning algorithms for age-related macular degeneration

Retinal imaging is important for diagnosis and monitoring of age-related macular degeneration (AMD), the main cause of severe vision loss in the elderly. However, interpretation of these images requires expert knowledge, is time-consuming and can be subjective. Deep learning is a form of machine learning based on deep neural networks that can be used to develop models that automatically classify or segment images. We investigate the applications of deep learning for automated analysis of retinal imaging for AMD.
In Chapter 3 we present a method for automatic detection of the fovea, the anatomical region in the center of the macula responsible for sharp central vision and perception of color. Chapter 4 describes a novel convolutional neural network architecture that is specifically designed for segmentation in a subset of the dimensions of the input data. We demonstrate its applicability on optical coherence tomography, where a 2-dimensional cross-sectional image represents information on a 1-dimensional line on the retinal surface. In Chapter 5 we describe a deep learning model for the automatic segmentation of geographic atrophy, a form of retinal pathology occurring in the late stage of AMD. We use this model to investigate growth of the atrophic area on a large data set with follow-up images. In Chapter 6 we present a model for segmentation of 13 features associated with neovascular and atrophic AMD. This model can automatically extract parameters such as total lesion volume, area or lesions count.
The performance of presented models is often comparable to expert human graders. We demonstrate their application in a research setting for better understanding AMD. Additionally these models could provide decision support when integrated in a clinical setting. Therefore deep learning could facilitate better treatment decisions and ultimately lead to better vision for people with AMD.