Kulyabin M (2025)
Publication Language: English
Publication Type: Thesis
Publication year: 2025
URI: https://open.fau.de/handle/openfau/36400
Retinal diseases impact the back layer of the eye, known as the retina, which is responsible for converting light into electrical signals transferred through the optic nerve to the brain. Eye diseases such as Age-related Macular Degeneration (AMD) and Diabetic Retinopathy (DR) often lead to progressive vision loss in the central visual field, affecting 170 million people worldwide. Artificial Intelligence (AI) can improve diagnostic quality in ophthalmic clinics, especially in the early stages. In this work, we presented several new methods to support medical experts in decision-making, raising a question about the decision-support system for ophthalmologists. Electroretinography (ERG) testing is a noninvasive clinical test that records the retina's electrical response to a brief flash of light as a waveform signal. The test measures the responses of various cell types in the retina, including the photoreceptors, inner retinal cells, and ganglion cells. Therefore, it is part of the standard testing procedure in ophthalmology clinics to detect retinal diseases such as AMD and DR. The ERG signal consists of two components: the a-wave and b-wave, and the magnitude of these values is typically used to detect abnormalities. Earlier, the same principle was used to apply machine learning models to classify ERG signals by training them on wave parameters such as amplitude and latency. However, this approach does not utilize potentially important information within the signal itself. In this thesis, we presented several works where we used Continuous Wavelet Transform (CWT) with different mother wavelet functions in combination with various Deep Learning (DL) models. The best result was obtained by a Vision Transformer (ViT) trained on a combination of three wavelets – Morlet, Complex Gaussian Derivative, and Ricker wavelets. On the collected dataset consisting of child and adult signals, we obtained balanced accuracies of 0.88, 0.85, and 0.91 for Maximum, Scotopic, and Photopic protocols, respectively. The ERG signal has other applications as a biomarker for various disorders. As the retina is an extension of the Central Nervous System (CNS), ERG provides access to CNS functionality. Earlier, ERG waveform analysis has shown potential as a new biomarker for Autism Spectrum Disorder (ASD), a neurodevelopmental condition that affects the CNS. Identifying it as early as possible is essential to provide the appropriate support to the individual and their family. In this thesis, we demonstrated a new method to detect ASD using DL models. A combination of ERG analysis in different domains and AI provides a novel method for detecting ASD in children at the earliest stages. On a collected dataset in Australia and England, consisting of ASD and control children, we trained several novel neural network architectures. We obtained the best result with the Gated Multilayer Perceptron (gMLP) model, with a balanced accuracy of 0.897. Thus, we reported that the ERG waveform has the potential to be a functional retinal biomarker and, in conjunction with gMLP, could further improve the accuracy of ASD detection. Optical Coherence Tomography (OCT) is an imaging technique that uses interferometry with short-coherence-length light. It is used in ophthalmology to visualize the retina at the back of the eye. OCT allows scanning through the retinal layers, providing images of their structure, and has become a standard tool for diagnosing and monitoring the treatment of retinal diseases. Recently, neural networks have been increasingly used to analyze, register, segment, and classify OCT scans. The essential point in the training of neural networks is data. OCT images are biometric; therefore, there is a lack of publicly available data. In this thesis, we presented a dataset of 2000 images from over 800 patients, divided into categories such as AMD, Diabetic Macular Edema (DME), Epiretinal Membrane (ERM), Retinal Artery Occlusion (RAO), Retinal Vein Occlusion (RVO), and Vitreomacular Interface Disease (VID). We trained the classical Convolutional Neural Network (CNN) models ResNet50 and VGG16 on our dataset, obtaining accuracies of 0.846 and 0.859, respectively. Publishing such datasets positively impacts the development of AI in ophthalmology and the design of decision-support systems. Retinal diseases, such as Stargardt’s disease, are effectively identified at the cellular level, particularly through the analysis of cone photoreceptors. Adaptive Optics Scanning Light Ophthalmoscopy (AOSLO) enables high-resolution, in vivo imaging of these cells by detecting light reflected from their structure. Although recent advances have improved automated algorithms for cone segmentation in confocal AOSLO images, annotating training data remains a largely manual and time-consuming task. In this thesis, we propose a novel approach that integrates a human-in-the-loop method with attention-augmented DL models and the Voronoi algorithm to improve the detection and segmentation of cones in AOSLO imagery. Trained on a labeled dataset, the method achieved F1 scores of up to 0.968 while reducing the workload of expert labeling.
APA:
Kulyabin, M. (2025). Artificial Intelligence for detection and control of retinal and neurodevelopmental disorders (Dissertation).
MLA:
Kulyabin, Mikhail. Artificial Intelligence for detection and control of retinal and neurodevelopmental disorders. Dissertation, 2025.
BibTeX: Download