Beyond the Eye: A Relational Model for Early Dementia Detection Using Retinal OCTA Images

TL;DR

The paper addresses the challenge of early dementia detection by leveraging retinal OCTA as a noninvasive proxy for brain pathology. It introduces PolarNet+, a polar-transform-based architecture that aligns OCTA regional analysis with the ETDRS grid, serializes 3D retinal features, and applies a rewiring graph module to capture inter-regional relationships. The approach provides regional importance maps and a regional-relationship graph, achieving state-of-the-art performance for EOAD and MCI detection across ROAD and ROMCI datasets while offering clinically interpretable explanations. This method has the potential to enable scalable, clinically friendly screening and deepen understanding of eye–brain biomarkers in neurodegenerative disease.

Abstract

Early detection of dementia, such as Alzheimer's disease (AD) or mild cognitive impairment (MCI), is essential to enable timely intervention and potential treatment. Accurate detection of AD/MCI is challenging due to the high complexity, cost, and often invasive nature of current diagnostic techniques, which limit their suitability for large-scale population screening. Given the shared embryological origins and physiological characteristics of the retina and brain, retinal imaging is emerging as a potentially rapid and cost-effective alternative for the identification of individuals with or at high risk of AD. In this paper, we present a novel PolarNet+ that uses retinal optical coherence tomography angiography (OCTA) to discriminate early-onset AD (EOAD) and MCI subjects from controls. Our method first maps OCTA images from Cartesian coordinates to polar coordinates, allowing approximate sub-region calculation to implement the clinician-friendly early treatment of diabetic retinopathy study (ETDRS) grid analysis. We then introduce a multi-view module to serialize and analyze the images along three dimensions for comprehensive, clinically useful information extraction. Finally, we abstract the sequence embedding into a graph, transforming the detection task into a general graph classification problem. A regional relationship module is applied after the multi-view module to excavate the relationship between the sub-regions. Such regional relationship analyses validate known eye-brain links and reveal new discriminative patterns.

Figures (6)

• Figure 1: Demonstration of two different retinal imaging modalities of the same eye. (a) The CFP image and (b) its corresponding macula-centered OCTA images, such as superficial vascular complex (SVC), deep vascular complex (DVC), and choriocapillaris (CC). (c) The ETDRS grid applied on the OCTA image: temporal-inner (TI), temporal-external (TE), superior-inner (SI), superior-external (SE), nasal-inner (NI), nasal-external (NE), inferior-inner (II), and inferior-external (IE) sectors indicated.
• Figure 2: A schematic illustration of the proposed PolarNet+ for EOAD/MCI detection using OCTA images and its regional-interaction analysis. (a)-(b) Illustrations of the polar and Cartesian coordinate systems, respectively. (c) The ETDRS grid in Fig. \ref{['fundus_octa_polar_trans_fig']}-(c) is applied to the OCTA image after polar transformation. (d) PolarNet+ categorizes the input OCTA images into EOAD, MCI, and healthy controls (HC). (e)-(f) Visualizations of importance maps and relationship graph. Different colors indicate different levels of significance. Annotations: ${O_p}$: the transformation center, also the FAZ center, also the pole of the polar coordinate system; ${O_c}$: the origin of the Cartesian coordinate system, where the horizontal axis is the ${X}$ axis, and the vertical axis is the ${Y}$ axis; ${R}$: the radius of the region of interest.
• Figure 3: The details of the proposed PolarNet+ (a) and its modules (b), (d) and (e). PolarNet+ comprises the spatial extension module, the multi-view module, and the regional relationship Module. PolarNet+ treats stacked images as volumes, and all operations are in 3D. We obtain the Sequencer3D block (d) by replacing the multi-head (MH) attention module in the transformer block (c) with BiLSTM3D (e). The right side of the subfigure (e) is the schematic representation of the 3D serialization processing. From top to bottom are explanations of the detailed splitting and mapping relations in the three directions (along the radius, around the pole, and along the depth). Here, we use one-way arrows for easy understanding. As for the implementation, we use BiLSTM for comprehensive feature extraction.
• Figure 4: Visualization of regional relationships by the means of regional adjacency matrix, generated from dataset ROAD-I.
• Figure 5: Visualization of regional importance maps (a), generated from dataset ROAD-I and ROMCI-I. For comparison, we generated (b) the results of the normal visualization method, Grad-CAM, in the AD detection task on ROAD-I from a ResNet-18. (c) and (d) are the outcomes of the regional statistical assessments of the parameters generated from dataset ROAD-I and ROMCI-I, adjusted for covariates including age, gender, hypertension, diabetes, and education level, which were acquired through the utilization of the generalized estimation equation.