Artificial Intelligence for Detection and Control of Retinal and Neurodevelopmental Disorders
The retina, the light-sensing layer at the back of the eye, converts light into electrical signals sent to the brain via the optic nerve. Diseases like Age-related Macular Degeneration (AMD) and Diabetic Retinopathy (DR) progressively damage central vision and affect roughly 170 million people worldwide. This thesis develops several AI-based methods to support ophthalmologists in diagnosing these conditions earlier and more reliably.
One standard diagnostic test, electroretinography (ERG), measures the retina’s electrical response to a flash of light. Traditionally, this signal is reduced to a few key measurements (amplitude and timing of its two main waves) before being fed into machine learning models, which throws away a lot of information contained in the raw waveform. This thesis instead applies Continuous Wavelet Transform, using several different wavelet functions, to extract richer features from the signal before classification. The best-performing setup, a Vision Transformer trained on a combination of three wavelet types, achieved balanced accuracies of 0.88, 0.85, and 0.91 across three different ERG testing protocols on a dataset of both children and adults.
Because the retina is effectively an extension of the central nervous system, ERG signals can also serve as a biomarker for neurodevelopmental conditions like Autism Spectrum Disorder (ASD), where early detection matters for getting timely support. Using ERG data from children in Australia and England, this thesis trains several neural network architectures to detect ASD, with a Gated Multilayer Perceptron achieving the best balanced accuracy of 0.897. This suggests ERG waveforms, combined with the right model, could become a useful functional biomarker for early ASD detection.
The thesis also addresses Optical Coherence Tomography (OCT), an imaging technique that lets clinicians see through the retina’s layers and has become standard for diagnosing and monitoring retinal disease. Since OCT images are biometric data, publicly available datasets are scarce, limiting the training of neural networks for this purpose. To help address this, the thesis contributes a dataset of 2,000 OCT images from over 800 patients across six disease categories, and trains standard CNN models (ResNet50 and VGG16) on it, reaching accuracies of 0.846 and 0.859 respectively.
Finally, the thesis tackles cone photoreceptor imaging, relevant to diseases like Stargardt’s disease that are best identified at the cellular level. Adaptive Optics Scanning Light Ophthalmoscopy (AOSLO) can image these cells directly, but labeling the necessary training data by hand is slow and labor-intensive. This thesis proposes a human-in-the-loop method combining attention-based deep learning with the Voronoi algorithm to detect and segment cones more efficiently, achieving F1 scores of up to 0.968 while cutting down the manual labeling workload.