AI to Identify Strain-Sensitive Regions of the Optic Nerve Head Linked to Functional Loss in Glaucoma

Abstract

Purpose

The purposes of this study were to assess whether optic nerve head (ONH) biomechanics, quantified by tissue strain, improves classification of progressive visual field (VF) loss patterns in glaucoma beyond morphology, and to use saliency maps to identify ONH regions associated with the predictions.

Methods

We recruited 249 patients with glaucoma (mean age 69 ± 5 years, 54% female patients). One eye per subject was imaged under (1) primary gaze and (2) primary gaze with IOP elevated to approximately 35 millimeters of mercury (mm Hg) via ophthalmo-dynamometry. Twelve subjects were excluded due to poor scan quality/limited lamina cribrosa (LC) visibility. Experts classified subjects into four categories based on the presence of specific visual field defects (VFDs): (1) superior nasal step (N = 26), (2) superior partial arcuate (N = 62), (3) full superior hemifield defect (N = 25), and (4) other/non-specific defects (N = 124). Automatic segmentation and digital volume correlation computed neural tissue and LC strains. Biomechanical and structural features were input to a PointNet model. Three classification tasks were performed to detect: (1) superior nasal step, (2) superior partial arcuate, and (3) full superior hemifield defect. Data were split 80/20 (train/test). Area under the curve (AUC) assessed performance. Saliency maps (an explainable artificial intelligence [XAI] technique) highlighted ONH regions most critical to classification.

Results

Models achieved AUCs of 0.77 to 0.88 across VFD classifications. The structure-only model reached an AUC of 0.83 ± 0.02 for superior arcuate defects, which significantly improved to 0.87 ± 0.02 (P < 0.05) with the addition of strain information, demonstrating that ONH biomechanics enhance prediction beyond morphology. Strain-sensitive regions were localized to the inferior and inferotemporal rim, expanding with increasing severity of VF loss.

Conclusions

ONH strain enhances classification of glaucomatous VF loss patterns. The neuroretinal rim, rather than the LC, was most critical, suggesting rim strain may play a dominant role in axonal injury and functional loss.

Our recent studies^1–6  have demonstrated a consistent association between intraocular pressure (IOP)-induced deformations in the optic nerve head (ONH) and patterns of visual field (VF) loss, supporting a biomechanical contribution to glaucoma pathogenesis, as echoed by others.^7–12 However, whereas this association is increasingly accepted, direct evidence that biomechanics drive functional decline remains limited, and its predictive value for VF progression is still uncertain.

A key challenge lies in identifying what biomechanical factors, such as stress, strain, or specific deformation modes, drive axonal injury, and where on the ONH such mechanical damage is most likely to occur. Some propose the retina as the initial site, others point to the lamina cribrosa (LC),¹³ and some suggest both.¹⁴ Clarifying the nature and location of these biomechanical triggers is essential to understanding how axonal degeneration unfolds in glaucoma and may inform more targeted approaches to patient risk assessment.

To address these questions, we strongly believe that recent artificial intelligence (AI) approaches, specifically geometric deep learning (to learn from complex 3D structural and biomechanical data)¹⁵ and explainable AI (XAI),¹⁶ offer powerful tools to investigate axonal injury mechanisms in glaucoma. XAI refers to methods that make AI model decisions interpretable by revealing which features or regions influence predictions; saliency maps are a common example that highlight areas most relevant to the model's output. These methods hold great potential to identify the specific types of mechanical stimuli and their precise locations that are most likely responsible for driving VF loss. Importantly, such techniques have already shown promise in glaucoma research by establishing critical links between ONH biomechanics and visual function.¹⁷

In this study, we aimed to: (1) assess whether ONH biomechanics improves the prediction of distinct patterns of VF loss that reflect progressive stages of glaucoma; and (2) use XAI to identify the ONH regions most associated with these predictions, providing insights into potential biomechanical pathways of axonal injury and disease progression.

Methods

Subject Recruitment

This study was approved by the SingHealth Centralized Institutional Review Board and adhered to the tenets of the Declaration of Helsinki. Written informed consent was obtained from each subject. This is a clinic-based cross-sectional study.

We recruited 249 subjects with glaucoma from clinics at the Singapore National Eye Centre. We included subjects aged more than 50 years old, of Chinese ethnicity (predominant in Singapore), with a refractive error within ±3 diopters, and excluded subjects who had undergone prior intraocular/orbital/brain surgeries, and subjects with a history of strabismus, ocular trauma, ocular motor palsies, and orbital/brain tumors. Subjects with clinically abnormal saccadic or pursuit eye movements, poor LC visibility on optical coherence tomography (OCT; <50% en face visibility),³ or OCT scans that failed automated segmentation by the REFLECTIVITY algorithm were also excluded. Additional exclusion criteria included known carotid or peripheral vascular disease, other abnormal ophthalmic or neurological conditions, visual field defects (VFDs) secondary to diabetic retinopathy or non-glaucomatous optic neuropathies, and advanced glaucoma (mean deviation < −21 decibel [dB]).

Glaucoma cases were defined by the presence of glaucomatous optic nerve damage, defined as vertical cup-disc ratio (VCDR) of >0.7 and/or neuroretinal rim narrowing (focal or diffuse thinning of the rim tissue, often presenting as notching or an increased vertical cup-to-disc ratio), with an associated VFD on standard automated perimetry. The latter was defined if the following were found: (1) glaucoma hemifield test outside normal limits, (2) a cluster of ≥3, non-edge, contiguous points on the pattern deviation plot, not crossing the horizontal meridian with a probability of <5% being present in age-matched normals (one of which was <1%), and (3) Pattern Standard Deviation (PSD) <0.05; these were repeatable on 2 consecutive occasions.

Visual Field Testing

Unreliable VF test results with a false-positive error of more than 15% or a fixation loss of more than 33% were excluded. The thresholding algorithm (Swedish interactive testing algorithm) standard 24-2 program was selected. For every selected patient, the PSD map from the Humphrey visual field report was used in the study. For each point in the PSD map, we defined a defect as a point having a P value of less than 5%.

Classification of Visual Field Defect Patterns

Our glaucoma expert (author M.N.) classified VFD patterns based on each patient’s VF report. The classification system used was consistent with our previous work,¹⁸ which was derived from the Ocular Hypertension Treatment Study classification system.¹⁹

Subjects were categorized into the following groups based on their VFD, excluding the first VF test: (1) presence of a superior nasal step; (2) presence of a superior arcuate defect; (3) presence of a full superior hemifield defect; and (4) other types of VFDs, including inferior or nonspecific defects. In eyes exhibiting overlapping defects (e.g. both a nasal step and an arcuate scotoma), classification was based on the predominant or more advanced pattern. Accordingly, when an arcuate defect extended from a nasal step region, the eye was classified as having a superior arcuate defect rather than a nasal step. This approach ensured mutually exclusive and progression-consistent categories.

We focused on superior VFDs, as they were the most prevalent in our dataset and are commonly observed in early glaucomatous damage.²⁰ These categories represent a spectrum of superior VF loss, ranging from mild (nasal step) to more advanced damage (full hemifield defect).

OCT Imaging and Biomechanical Testing

We selected one eye at random from each patient with bilateral glaucoma, and we imaged the ONH with spectral-domain OCT (Spectralis; Heidelberg Engineering GmbH). The imaging protocol was similar to that from our previous work.³ In brief, we conducted a raster scan of the ONH (covering a rectangular region of 150 × 100 degrees centered at the ONH) comprising 97 serial B-scans, with each B-scan comprising 384 A-scans. The average distance between B-scans was 35.1 µm, and the average axial and lateral A-scan resolutions were 3.87 µm and 11.5 µm, respectively. Each eye was scanned twice under two conditions: OCT at baseline and under acute IOP elevation.

Specifically, for the study eye of each subject, the IOP was measured using a Tonopen post pupillary dilation with tropicamide 1% (Alcon, Puurs, Belgium) and an OCT volume of the ONH was acquired at baseline IOP. The IOP was then raised by gentle indentation of the temporal side of the lower eyelid (perpendicular to the sclera), using an ophthalmodynamometer (ODM). After the ODM application, IOP was held constant and was measured with a Tonopen while the indenter was maintained in place to acquire the second OCT volume under elevated pressure. The forces for IOP increase were chosen to correspond to final IOP values of approximately 35 millimeters of mercury (mm Hg; based on a preliminary assessment performed on 20 healthy eyes).²¹ The applied force was consistently 0.65 N, as calibrated using a uniaxial tensile tester (Instron-5848; Instron, Inc., Noorwood, MA, USA).

AI-Based Segmentation of ONH Tissues and Representation of the ONH Structure as 3D Point Cloud

We automatically segmented all baseline OCT volume scans of the ONH using REFLECTIVITY (Reflectivity; Abyss Processing Pte Ltd., Singapore). More specifically, the following ONH tissue groups were automatically labeled (Fig. 1a): (1) the retinal nerve fiber layer (RNFL) and the prelamina tissue; (2) the ganglion cell inner plexiform layer; (3) all other retinal layers; (4) the retinal pigment epithelium with Bruch's membrane (BM) and the BM opening (BMO) points; and (5) the OCT-visible part of the LC. In almost all OCT volume scans, the posterior boundaries of the LC were not visible and could not be segmented.

Figure 1.

(a) Segmentation of ONH neural tissues consisting of the retinal nerve fiber layer (RNFL) and the prelamina tissue (PLT), ganglion cell inner plexiform layer (GCL + IPL), all other retinal layers (ORL), the retinal pigment epithelium (RPE) and the lamina cribrosa (LC). (b) An illustration of ONH point cloud containing both the tissue thickness and the effective strain as an input to the PointNet. The output is the binary classification of three separate classification tasks. (c) An example of a saliency map representing the importance of each region to each classification task.

Each ONH was then represented as a 3D point cloud derived from the segmentation of OCT images and served as input to our geometric deep learning algorithm, following an approach established in our previous work.³ All ONH point clouds were rigidly aligned with respect to the BMO center and plane to ensure consistent anatomic orientation across subjects. Full methodological details are provided in the Supplementary Appendix A.^22–24

Deformation Tracking and IOP-Induced Effective Strain Derivation

To extract the 3D deformation of the ONHs under an acute elevation in IOP, we used a commercial digital volume correlation (DVC) module (Amira version: 2020.3; Thermo Fisher Scientific, USA). Details about the preprocessing of the volumes and DVC algorithm can be found in our previous work.³ From the deformation mapping, we computed the effective strain within the ONH, a measure of local 3D deformation that accounts for both compressive and tensile components with higher values indicating greater mechanical deformations in either direction. We then assigned effective strain values to each point in the ONH point cloud using K-nearest neighbor interpolation (K = 5; Fig. 1b).

Classification of Visual Field Defects Using Geometric Deep Learning

In this study, we used PointNet to take each ONH point cloud (see example in Fig. 1b) as an input to predict the corresponding visual field defect map (52 points) for that specific eye.²³

We trained three separate PointNet models, each targeting a distinct classification task:

• (1)Detecting the presence of a superior nasal defect (presence: N = 26; absence: N = 211);
• (2)Detecting the presence of a superior arcuate defect (presence: N = 62; absence: N = 175);
• (3)Detecting the presence of a full superior hemifield defect (presence: N = 25; absence: N = 212).

Each model was trained to classify the presence or absence of a specific defect pattern based on 3D ONH morphology and biomechanical strain. The network was trained using binary cross-entropy loss. Model hyperparameters were tuned based on our previous work.¹⁷ To approximate the model's performance, we performed a stratified five-fold cross-validation and report the mean areas under the curve (AUCs) ± standard deviations with confidence interval (CI).

To evaluate the impact of different feature sets—namely, all features, ONH point clouds with thickness, ONH point clouds with strain, and ONH point clouds alone—we assessed model performance in classifying the superior arcuate pattern by performing a stratified 5-fold cross-validation, reporting the mean AUCs ± standard deviation.

During the training process, we augmented the point clouds in each sample using various techniques: random cropping, random rotations, and random sampling of points (selecting a subset of 3000 points from each point cloud).²⁵ This augmentation enhances the model’s generalizability to unseen datasets.

Saliency Maps Identify Optic Nerve Head Regions Associated With the Prediction of Visual Field Defect Types

To identify the ONH regions most critical for prediction, we used gradient-based saliency maps, an XAI technique that highlights regions where structural and strain features most influence the model's output.²⁶ Gradient magnitudes across spatial, structural, and strain features were averaged to reflect each point's importance. Because saliency maps are represented in 3D, we simplified their visualization. Saliency values (gradients) were sum-projected onto an en face plane (2 × 2 mm grid centered on BMO, with 40 × 40 cells) and also visualized along inferior-superior and nasal-temporal ONH cross-sections. Note that all ONHs were scaled relative to the BMO radius to ensure consistent anatomic representation across subjects, and all saliency maps were normalized to a scale between 0 and 1. Saliency maps were generated for each of the three classification tasks: detection of nasal steps, arcuate defects, and hemifield defects.

Finally, from the saliency maps, we defined regions of high gradient as sensitive regions, area where local strain and tissue thickness most strongly contribute to the prediction of characteristic visual field loss patterns in glaucoma.

Importance of Biomechanical Strain in Predicting Visual Field Loss

To assess the contribution of biomechanical strain features to model performance, we conducted a sensitivity analysis focused on the classification of superior arcuate visual field defects. This pattern was selected because it represents a well-characterized and relatively common form of early glaucomatous damage,^20,27 and, critically, was the only superior defect subtype in our dataset with a sufficiently balanced sample size (presence: N = 62 and absence: N = 175). In contrast, the superior nasal step (N = 26) and full superior hemifield defect (N = 25) groups had markedly fewer cases.

We trained the model with and without strain features and evaluated performance across five independent train-test splits. Results were reported as the AUC, expressed as mean ± standard deviation.

Comparison of Principal Strain Components in Predicting Visual Field Loss

To evaluate the contribution of different strain modes—maximum tensile, maximum compressive, and maximum shear strain—we performed an additional analysis focused on classifying superior arcuate visual field defects. In this experiment, each strain component was independently substituted for effective strain in the model, together with the same set of morphological parameters. Model performance for each strain type was assessed using five-fold cross-validation, and the mean AUC ± standard deviation was reported for comparison.

Evaluation of Potential Optical Distortions Introduced by ODM Use

In principle, the use of an ODM could introduce optical distortions by altering corneal curvature or lens geometry. To assess this possibility, we measured central corneal thickness and anterior chamber depth in 55 eyes, both with and without ODM application. We also assessed corneal curvature (R1 and R2) using a NIDEC keratometer. We found no significant differences in central corneal thickness (baseline = 550 ± 34 µm vs. ODM = 549 ± 35 µm) or anterior chamber depth (baseline = 3.6 ± 0.8 mm vs. ODM = 3.5 ± 0.8 mm). Corneal curvature measurements in a subject similarly revealed minimal changes in R1 (baseline = 7.96 mm vs. ODM = 7.95 mm) and R2 (baseline = 7.74 mm in both cases). These results indicate that any optical distortion induced by ODM use is minimal and unlikely to affect our measurements. To determine whether ODM use affected OCT image quality, we compared signal-to-noise ratio (SNR) between baseline and ODM scans in five eyes. The mean change in SNR (ODM – baseline) was 0.0 ± 1.6 dB, with no significant difference detected by a paired t-test (P = 0.9).

Results

Subjects’ Demographics

A total of 249 Chinese subjects with glaucoma were initially recruited for this study; however, 12 were excluded due to limited LC visibility and/or poor scan quality, resulting in 237 subjects included in the final analysis. The analyzed cohort had a mean age of 69 ± 5 years, and 128 (54%) were women. Grouping into visual-field defect categories was performed after recruitment. The cohort spanned a broad spectrum of disease severity, with mean deviation (MD) values ranging from −1.8 dB (mild) to −20  dB (severe), and an average MD of −7.25 ± 5.05 dB. Among these subjects, 26 presented with a nasal step defect, 62 had a superior partial arcuate defect, 25 exhibited a full superior hemifield defect, and the remaining 124 had other or nonspecific visual field abnormalities. A detailed summary of demographic and clinical characteristics stratified by visual field defect subtype is provided in the Table.

Table.

Demographic and Clinical Characteristics of Subjects Across Visual Field Defect Categories

Characteristic, Mean ± Standard Deviation	Superior Nasal Step Defect (N = 26)	Superior Arcuate Defect (N = 62)	Superior Hemifield Defect (N = 25)	Nonspecified or Other Defects (N = 124)	P Value
Age, y	69 ± 6	69 ± 6	70 ± 4	70 ± 5	0.16
Sex, female	79%	54%	52%	64%	0.04 *
Axial length, mm	24.0 ± 1.0	24.0 ± 1.0	23.8 ± 0.8	23.7 ± 1.0	0.04 *
Visual field, MD, dB	−7.5 ± 4.6	−8.0 ± 4.4	−14.0 ± 3.6	−9.4 ± 5.3	<0.0001 *
Pattern standard deviation, dB	7.9 ± 4.9	9.0 ± 4.0	13.0 ± 2.7	9.7 ± 5.5	<0.0001 *
Baseline IOP, mm Hg	15.8 ± 3.4	15.8 ± 3.0	14.7 ± 2.7	15.0 ± 3.7	0.17
IOP, mm Hg with indentation	35.9 ± 4.4	33.8 ± 6.4	35.0 ± 5.5	34.8 ± 5.9	0.21

Values are presented as mean ± standard deviation or as percentages where indicated. P values correspond to overall group comparisons obtained using 1-way ANOVA for continuous variables and the chi-square test for categorical variables (sex).

^*Denote statistically significant differences (P < 0.05).

Model Performance Across Visual Field Defect Classifications

The geometric deep learning model achieved consistently strong performance across all classification tasks, with mean AUC ± SD (95% CI) across 5 independent folds of 0.77 ± 0.04 (0.74–0.80) for nasal step detection, 0.87 ± 0.02 (0.85–0.88) for superior partial arcuate defects, and 0.85 ± 0.03 (0.83–0.88) for superior hemifield defects (Fig. 2).

Figure 2.

Comparison of visual field defect patterns by severity. (a) Nasal step, (b) superior arcuate, and (c) superior hemifield loss are illustrated with corresponding visual field defect samples from our population. Red regions in the nerve fiber defect maps denote areas of nerve loss, aligned with established structure-function relationships such as the Garway-Heath map. (d–f) Saliency maps highlight an arching pattern, where the arc length increases with defect severity. Red regions indicate areas of high correlation with each defect pattern. To provide anatomical context, a sample ONH is overlaid in the background, referencing the neuroretinal rim. Additional cross-sectional views are provided: one along the superior-inferior axis through the ONH’s midsection, and another along the nasal-temporal direction at the neuroretinal rim.

Incorporation of Effective Strains Improves Prediction of Superior Arcuate Pattern

The model incorporating biomechanical strain significantly outperformed the one without it in predicting superior arcuate defects (AUC = 0.87 ± 0.02 vs. 0.83 ± 0.02, P < 0.05), highlighting strain as a meaningful predictor of region-specific visual field loss. Among input configurations, the full feature set yielded the highest performance, followed by point clouds with thickness (AUC = 0.83 ± 0.02), strain (AUC = 0.78 ± 0.01), and point clouds alone (AUC = 0.75 ± 0.02). These findings suggest that biomechanical strain provides complementary information to structural metrics and may reflect underlying mechanisms of glaucomatous damage.

Saliency Maps Reveals an Arching Pattern in the Infero-Temporal Neuroretinal Rim, Increasing With Glaucoma Severity

Across all models, critical regions from the saliency maps (also defined as strain-sensitive regions) consistently appeared in the inferior and infero-temporal neuroretinal rim (see Fig. 2). In the en face view, a spreading pattern was observed, extending from the inferior to the temporal region (Figs. 2d, 2e), forming an arc that increased in length as the defect progressed from a nasal step to a arcuate pattern and, ultimately, to full hemifield loss (all superior). As the arc expanded, it covered an increasingly larger proportion of axons entering the disc. Additionally, a less prominent critical region was observed in the superior neuroretinal rim, which may reflect inferior visual field defects in some patients.

Although the primary analyses focused on superior defects, additional analyses were performed on eyes with inferior arcuate patterns and compared with inferior defects; these results are presented in Supplementary Appendix B.

Evaluation of Tensile, Compressive, and Shear Strain Components for Visual Field Prediction

We found that the use of compressive strains yielded the highest AUC (0.85 ± 0.02), followed by maximum shear strains (0.84 ± 0.02), and maximum tensile strain (0.82 ± 0.02). However, effective strain (a combination of all components) yielded the overall best performance (0.87 ± 0.02).

Discussion

In this study, we demonstrated that incorporating ONH tissue strain into AI models significantly improved the prediction of visual field loss patterns in glaucoma beyond what was possible using structural features alone. Using geometric deep learning on 3D point clouds of the ONH and acute IOP-induced strain measurements, we achieved high classification performance across 3 visual field defect types relevant to glaucoma progression. Importantly, saliency maps consistently highlighted the inferior and infero-temporal neuroretinal rim as the most critical region contributing to these predictions. These findings raise the possibility that local IOP-induced strain in the neuroretinal rim is associated with axonal injury and may help explain patterns of region-specific visual field loss in glaucoma.

In this work, we found that incorporating biomechanical strain information significantly improved the prediction of a localized visual field defect. Unlike our earlier works, which demonstrated that strain enhances prediction of full-field visual sensitivity maps and revealed a negative association between ONH strain measures and MD values,^3,17 the current approach targeted superior visual field defect, as this is the more commonly observed in glaucoma.²⁸ Specifically, we were able to reveal a more spatially resolved link between IOP-induced mechanical deformation and site-specific visual loss. This localized relationship is biologically plausible, as strain has been suggested to (1) cause direct mechanical injury to retinal ganglion cells (RGCs),^29,30 (2) disrupt microcapillary blood flow in the lamina cribrosa, choroid, and retina,^31,32 and (3) impair axoplasmic transport.³³ These deformation-driven pathways compromise axonal integrity and contribute to vision loss in glaucoma. Our findings suggest that local biomechanical strain is not only a global predictor of visual function but also provides valuable insight into the region-specific vulnerability of the ONH.

We found that strain-sensitive regions of the ONHs increased in size with glaucoma severity and were consistently concentrated around the inferior and infero-temporal neuroretinal rim, consistent with structure-function relationship in eyes with superior VFD. Specifically, for eyes with superior nasal defects, our AI models consistently identified the inferior neuroretinal rim as the most critical region. With worsening VF loss from superior arcuate to full superior hemifield defects, these critical regions expanded to include the inferotemporal and inferonasal rims. One way to interpret these findings is that, as glaucoma progresses, strain affecting an increasing proportion of axons entering the optic disc becomes predictive of expanding VFDs, which could be consistent with a developing pattern of axonal injury. Confirming this hypothesis will require longitudinal monitoring of patients undergoing biomechanical testing, which is currently underway in our laboratory.

Importantly, these transitions were entirely data-driven and emerged naturally from the model, without reliance on any predefined anatomic maps. Yet, the spatial relevance patterns identified by our AI models align well with established structure-function relationships. For instance, they correspond with the Garway-Heath map,³⁴ which links superior nasal field loss to damage in the inferior rim and associates more extensive VFDs with broader regions of the optic disc. Unlike the Garway-Heath framework, however, our method does not rely on pre-defined sectors. Instead, it uses geometric deep learning and patient-specific saliency maps that capture biomechanics-function associations – offering a potential key to uncovering the mechanisms of axonal injury.

Another notable observation from our saliency map (see Fig. 2) is the lower importance assigned to the LC region compared to the prelaminar tissue. This is intriguing, as the LC undergoes significant remodeling in glaucoma, such as excavation and increased curvature.^35–37 Additionally, the LC is a key region where biomechanical strain accumulates,³⁸ a factor that has been linked to glaucoma progression in both our study and previous research.^3,39 Given these, we would expect the LC to have greater importance, particularly in cases of more severe glaucomatous damage, such as hemifield defects. This discrepancy may arise from several factors, including limitations in how the model interprets the LC's role, constraints in feature representation learning within PointNet,⁴⁰ insufficient resolution of LC pores using current clinical OCT technology,⁴¹ or the small number of positive samples for certain defect types in our dataset. In addition, incomplete LC visualization in many eyes, where only the anterior portion was visible on OCT and portions were obscured by major vessel shadows, likely reduced the apparent contribution of the LC in our predictions. Alternatively, it may reflect a true biological sensitivity, as axons in the neuroretinal rim must undergo a sharp 90-degree turn – a potential point of mechanical vulnerability.⁴² Advances in OCT imaging may help clarify these possibilities.

An interesting observation from our analysis is that the model assigned greater importance to the nasal-inferior region in eyes with superior hemifield defects, forming an arcing pattern that extends temporally. Clinically, such defects are more often associated with structural loss in the inferior-temporal ONH region,⁴³ however, this difference may reflect distinct aspects of tissue biomechanics versus morphology. Strain-based features capture regions of mechanical susceptibility rather than visible thinning, and the combined strain-thickness representation may highlight areas where mechanical and structural variability jointly contribute to the model's prediction. The smaller number of eyes with superior hemifield defects compared with partial arcuate defects (25 vs. 62) may also have influenced the observed spatial distribution. Notably, in partial arcuate defects, the saliency pattern localized as expected to the inferior-temporal region. Ongoing recruitment and longitudinal studies, together with new analytic approaches to disentangle the roles of strain and thickness, will further clarify whether the nasal emphasis represents an early or complementary biomechanical response within the broader structure-function cascade.

We also conducted a preliminary analysis comparing individual strain components and found that models based on maximum compressive, tensile, or shear strain achieved slightly lower average AUCs (0.85 ± 0.02, 0.82 ± 0.02, and 0.84 ± 0.02, respectively) than those using effective strain (AUC = 0.87). These findings suggest that the aggregate strain measure captures a balanced representation of the dominant mechanical effects contributing to VF loss. Interestingly, our observations are consistent with Sigal et al.,⁴⁴ who reported that axons within the lamina cribrosa and post-laminar regions predominantly experience longitudinal compression and transverse stretch, whereas those in the prelaminar and rim regions undergo longitudinal stretch and transverse compression. Future tensor-based models incorporating full strain fields, combined with larger longitudinal datasets, will enable a more detailed and directional understanding of these biomechanical patterns and their role in glaucomatous axonal injury.

Several aspects of this study warrant further discussion. First, although we use gradient-based saliency maps as an XAI tool to visualize ONH regions that most influence the model's predictions, these maps provide associative insight and cannot, on their own, determine whether a highlighted region reflects the presence of a feature, the absence of a feature, or correlated inputs. In addition, our PointNet architecture aggregates features early, which limits the ability to disentangle regional contributions of strain versus tissue thickness within the saliency map. To address these XAI limitations, we are developing methods that separately assess the influence of strain, morphological, and vascular features (e.g. structured perturbations/ablation and attribution with grouped inputs) to quantify their independent and interactive effects.

Second, the subjects were recruited from a tertiary glaucoma clinic, and the cohort represents a selected population with a potentially more severe disease spectrum than the general population. Thus, these factors may affect the generalizability of our findings and underscores the need for a larger, more diverse dataset in future studies. Moreover, the VFD types we examined, superior nasal step, superior arcuate, and superior hemifield loss, do not preclude the presence of additional defects in the inferior region, which may explain the less prominent critical regions observed in the superior neuroretinal rim (see Figs. 2d–f). As shown in the Supplementary Appendix B, our additional analysis revealed that inferior arcuate defects exhibit a pattern opposite to that of superior arcuate defects. However, we did not include further examples of inferior arcuate defects due to the limited sample size. Additionally, this study did not include healthy controls, as our objective was to learn local biomechanics-function relationships within the glaucomatous disease space. Healthy eyes do not exhibit VFDs and therefore provide no signal for this specific task. Future work will leverage our previously published datasets in healthy cohorts,^2,45 as well as a new healthy cohort currently under recruitment at Emory, to characterize local biomechanical differences between healthy and glaucomatous eyes.

Third, the relatively small and imbalanced subgroup sizes may limit model generalizability. Although cross-validation and confidence bounds provide approximate performance estimates, independent validation is still needed to confirm robustness. Moreover, the cross-sectional design prevents evaluation of temporal or predictive effects. Despite these limitations, this study serves as a proof of concept, showing that geometric deep learning can produce interpretable saliency maps highlighting biomechanically relevant regions associated with VF loss.

Fourth, strain was derived from pairwise registration between baseline and IOP-elevated scans, capturing each eye's relative mechanical response to an acute IOP increase. This pressure change (ΔIOP) varied slightly across subjects depending on their baseline and elevated IOP levels. In exploratory analyses, we normalized strain by ΔIOP, as in our previous work, and found no appreciable change in model performance, likely because baseline and elevated IOP values were tightly clustered across groups. For this reason, we report un-normalized strain values in the main analysis.

Fifth, although the use of an ODM could theoretically introduce optical distortion by altering corneal geometry, our results indicate that such effects are negligible. Measurements of anterior segment parameters and corneal curvature showed no meaningful differences with ODM application, and OCT image quality, as assessed by SNR, was unchanged. Together, these findings suggest that ODM use does not introduce clinically relevant optical distortion or degrade OCT image quality, and therefore is unlikely to confound the measurements or conclusions of the present study.

Last, because of the cross-sectional nature of this study, the associations between ONH strain and VF loss should be interpreted as correlative rather than causal. It remains unclear whether elevated strain represents a precursor to axonal injury or a consequence of tissue remodeling in already-damaged regions. Longitudinal studies are therefore essential to establish the temporal sequence of biomechanical changes and to substantiate the proposed mechanobiological cascade underlying glaucomatous progression.

In conclusion, our findings demonstrate that ONH tissue strain enhances the prediction of distinct VF loss patterns typically observed in glaucoma. Saliency maps highlighted the neuroretinal rim as the principal region influencing model predictions, whereas contributions from the LC should be interpreted cautiously given the limited visibility of the LC in our data. These results suggest that strain at the rim may represent a key biomechanical correlate of functional loss, whereas the role of deeper structures such as the LC warrants further investigation with higher-resolution imaging.