Early Alzheimer's Disease Detection from Retinal OCT Images: A UK Biobank Study

TL;DR

This work addresses the challenge of early Alzheimer’s disease detection using retinal OCT by applying end-to-end deep learning to raw OCT B-scans from the UK Biobank. It evaluates CNN and transformer architectures with an anatomically guided, multichannel OCT input and a year-weighted loss within a nested cross-validation framework, achieving a top mean AUC of $0.624$ ($\,pm 0.060$) for predictions up to $4$ years before diagnosis and $0.652 \,pm 0.058$ in an independent replication. Key findings include the superiority of ResNet-34 over VGG and OCT-specific transformers in this low-data setting, and ablation studies showing that combining raw OCT, layer-masked, and contour inputs yields the strongest performance; Grad-CAM analyses localize attention to the central macula, notably the BMEIS and IS/OSJ layers. The results provide a reproducible baseline for OCT-based preclinical AD prediction and highlight the need for larger, multi-center, multimodal datasets and external validation to translate OCT biomarkers into clinical screening tools.

Abstract

Alterations in retinal layer thickness, measurable using Optical Coherence Tomography (OCT), have been associated with neurodegenerative diseases such as Alzheimer's disease (AD). While previous studies have mainly focused on segmented layer thickness measurements, this study explored the direct classification of OCT B-scan images for the early detection of AD. To our knowledge, this is the first application of deep learning to raw OCT B-scans for AD prediction in the literature. Unlike conventional medical image classification tasks, early detection is more challenging than diagnosis because imaging precedes clinical diagnosis by several years. We fine-tuned and evaluated multiple pretrained models, including ImageNet-based networks and the OCT-specific RETFound transformer, using subject-level cross-validation datasets matched for age, sex, and imaging instances from the UK Biobank cohort. To reduce overfitting in this small, high-dimensional dataset, both standard and OCT-specific augmentation techniques were applied, along with a year-weighted loss function that prioritized cases diagnosed within four years of imaging. ResNet-34 produced the most stable results, achieving an AUC of 0.62 in the 4-year cohort. Although below the threshold for clinical application, our explainability analyses confirmed localized structural differences in the central macular subfield between the AD and control groups. These findings provide a baseline for OCT-based AD prediction, highlight the challenges of detecting subtle retinal biomarkers years before AD diagnosis, and point to the need for larger datasets and multimodal approaches.

Figures (3)

• Figure 1: Overview of the preprocessing pipeline and retinal layer annotations in OCT B-scans. (a) Original grayscale OCT scan with inner and outer retinal boundaries. (b) Retinal layer segmentation with 11 color-coded contours representing the anatomical boundaries. (c) Alternate view of the same segmented B-scan for visualization consistency. (d) Pixel-wise retinal layer mask used as input for the second channel of the model. The legend on the right maps each color to a specific retinal layer, from the inner limiting membrane (ILM) to the outer boundary of the retinal pigment epithelium (OB_RPE). This multichannel representation encodes both intensity and anatomical structure, providing richer input for deep learning models. Images used with permission from the UK Biobank under Application Number 82266.
• Figure 2: Composite RGB representation of a single OCT B-scan used as model input. (a) First channel: original grayscale OCT image. (b) The second channel: the layer-masked version, where the retinal layers are selectively enhanced to highlight the structural features. (c) Third channel: binary retinal layer contours providing anatomical boundary information. Images used with permission from the UK Biobank under Application Number 82266.
• Figure 3: Model interpretability analysis. (\ref{['fig:exp_regions']}) The aggregated top 5% most salient pixels for the AD (left, red/yellow) and CN (right, blue/green) classes highlight the model's consistent focus on distinct anatomical regions. (\ref{['fig:exp_examples']}) Individual examples show model attention for True Positives (TP), True Negatives (TN), False Negatives (FN), and False Positives (FP). Images used with permission from the UK Biobank under Application Number 82266.