Glaucoma, a leading cause of irreversible blindness, causes retinal ganglion cell death and disrupts visual information transmission through the optic nerve to the lateral geniculate nucleus (LGN). The optic radiations (OR) further transmit this information to the visual cortex. A key question is whether peripheral sensory changes affect central processing pathways. This study uses diffusion MRI (dMRI), a non-invasive method to assess white matter pathway properties, to investigate the relationship between glaucoma and alterations in OR tissue properties. Previous studies using dMRI in glaucoma have yielded inconsistent results, with some showing changes not specific to visual pathways. This study aims to address these inconsistencies by leveraging the large sample size of the UK Biobank (UKBB) dataset and advanced machine learning techniques to detect subtle and potentially non-linear patterns in the data. The UKBB dataset offers an unparalleled opportunity to mitigate confounding effects through statistical matching, creating a more robust analysis of the relationship between glaucoma and white matter changes in the OR.

Prior research exploring the effects of glaucoma on white matter using dMRI has produced mixed and sometimes contradictory findings. Some studies have reported changes not limited to the visual pathways, suggesting widespread reorganization or systemic effects beyond the visual system. Automated Fiber Quantification (AFQ) and diffusional kurtosis imaging (DKI) have been employed to quantify tissue properties from dMRI data, providing measures such as mean diffusivity (MD), fractional anisotropy (FA), and mean kurtosis (MK). These measures are sensitive to biological changes, but their interpretation can be complex and require careful consideration of potential confounders. Machine learning, particularly convolutional neural networks (CNNs), are increasingly used to analyze high-dimensional, non-linear data like dMRI tract profiles, offering the potential to improve the accuracy and sensitivity of detecting subtle disease-related changes compared to traditional statistical methods.

Data from the UK Biobank, including dMRI and health data from 905 glaucoma patients and 5292 healthy controls, was utilized. Statistical matching was employed to create a matched dataset (Dataset A) of 856 glaucoma patients and 856 controls, carefully matched for age, sex, ethnicity, and socioeconomic status, to mitigate confounding factors. Dataset A was divided into training (64%), testing (20%), and validation (16%) sets. Two additional test datasets (Datasets A.1 and A.2) were created for assessing generalization: Dataset A.1 for age-related macular degeneration (AMD) and Dataset A.2 for age-group classification. A separate dataset (Dataset B) was created for age-group classification (70 vs. 60 years old), with corresponding generalization datasets (B.1 and B.2). Automated fiber quantification (AFQ) using a Python-based pipeline (pyAFQ) was used to delineate the optic radiations (OR), corticospinal tract (CST), and uncinate fasciculus (UNC) in each subject's dMRI data. The diffusional kurtosis model (DKI) was used to compute tract profiles (FA, MD, MK) at 100 points along each bundle, with the first and last 10 nodes excluded to reduce partial volume effects. One-dimensional convolutional residual neural networks (CNNs) and L2-regularized logistic regression models were trained on the tract profiles to predict glaucoma status. Model performance was evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. Statistical significance of AUCs was assessed using DeLong's test and Bonferroni or FDR correction for multiple comparisons.

In the matched Dataset A, CNNs trained on OR tissue properties showed significantly higher accuracy (AUC = 0.69) in classifying glaucoma compared to CNNs trained on CST (AUC = 0.57) and UNC (AUC = 0.53). The OR CNN also outperformed the logistic regression model trained on the same data (AUC = 0.63, p = 0.0303). Importantly, the glaucoma-trained CNN did not generalize to AMD or age classification tasks (Datasets A.1 and A.2), indicating that the identified features are specific to glaucoma and not simply reflective of accelerated aging. Similar findings were observed in Dataset B, which was used to train age-group classification models. While the CNN trained to classify age showed similar AUCs across bundles (OR, CST, UNC), it failed to significantly distinguish between glaucoma and controls in datasets B.1 and B.2. These findings collectively suggest that the differences observed in OR tissue properties are unique to glaucoma and are not a simple consequence of normal aging or other retinal diseases.

The study's findings demonstrate a distinct and potentially non-linear signature of glaucoma in the tissue properties of the optic radiations. The superior performance of the CNN compared to the linear model suggests a complex, non-linear relationship between glaucoma and OR microstructure. The lack of generalization to age and AMD suggests that the observed changes are not simply a reflection of accelerated aging or general retinal dysfunction but are rather specific to the pathological processes of glaucoma. Several potential explanations for these findings are discussed, including visual input alterations leading to neural reorganization, transsynaptic degeneration through the LGN, and the possibility that the algorithm is specifically detecting a subgroup of glaucoma patients with distinct OR changes. The correlational nature of the study necessitates further investigation to fully establish the causal relationships involved.

This study provides evidence of a specific, non-linear signature of glaucoma in optic radiation tissue properties detectable by convolutional neural networks. This signature does not appear to be simply related to age or other retinal disorders. Future research should investigate the underlying biological mechanisms contributing to these alterations and explore the potential of these findings for diagnostic and prognostic applications. Further studies using more diverse and larger datasets are necessary to confirm and extend these findings.

The study's reliance on self-reported glaucoma status may introduce some classification errors. Variations in glaucoma sub-type, severity, duration, and treatment could also affect the results. The study is correlational and does not establish causality. Further limitations include the challenges in detecting the OR due to its high curvature and the use of phenomenological models for tissue property estimation (DKI).