Seminar - Yasemin Turkan

Dr. YASEMIN TURKAN will give a seminar on Deep Learning-Based Analysis of Retinal OCT Scans for Early Detection of Alzheimer's Disease from UK Biobank Dataset on the 11th of June Thursday, 11.00, in room A312. You are cordially invited. 

Deep Learning-Based Analysis of Retinal OCT Scans for Early Detection of Alzheimer's Disease from UK Biobank Dataset
Alzheimer's disease (AD) is the most common form of dementia. It is an irreversible, progressive brain disorder marked by a decline in cognitive functioning with no treatment. It is characterized by a massive decrease in brain size due to the accumulation of proteins (amyloid-beta and tau) in neurons. The eyes extend the brain, as both the retina and the brain develop from the same neural tube. Postmortem studies in AD also highlighted the collection of these proteins in the retina. More recently, high-resolution visual imaging techniques, including optical coherence tomography (OCT), have been proposed as tools for evaluating structural changes in the retina of AD patients. This research investigates the potential of retinal Optical Coherence Tomography (OCT) as a non-invasive, cost-effective biomarker for the early prediction of AD, utilizing the large-scale longitudinal data from the UK Biobank. A significant limitation in existing OCT analysis is the reliance on isolated 2D B-scans or computationally prohibitive 3D volumetric modeling. To overcome these challenges, we proposed an anatomically guided framework that utilizes 3D-informed en-face thickness projection maps. Based on preliminary interpretability studies and saliency mapping, our pipeline was optimized to focus on the 3mm. inner macular region, filtering out peripheral noise to prioritize diagnostically relevant features. Our results identify the Ganglion Cell Layer (GCL) as the most potent morphological indicator of preclinical AD. GCL thickness maps achieved a peak Mean AUC of 0.750 ± 0.037. Notably, the Retinal Nerve Fiber Layer (RNFL)(a traditional clinical benchmark) exhibited negligible predictive value in this pre-symptomatic cohort, remaining near the 0.5 chance baseline. Longitudinal sensitivity analysis further established a "diagnostic horizon" for retinal biomarkers. We observed that predictive accuracy is highest between 1 and 8 years before clinical diagnosis, with signals progressively converging toward baseline by the 12-year mark. When benchmarked against current literature, our structural-only GCL framework outperformed existing baselines for symptomatic Mild Cognitive Impairment (MCI) diagnosis, demonstrating its robustness in the much more challenging task of preclinical prediction. This research provides a reproducible baseline for population-level AD screening. By demonstrating that 3D-informed 2D representations of the GCL can effectively capture early neurodegenerative signals.