This study was approved by the institutional review board of Samsung Medical Center (SMC; 2024-05-131) and followed the principles of the Declaration of Helsinki. All data were collected retrospectively from the institution’s data management team, and all direct identifiers, such as names and resident registration numbers, were fully deidentified through a standardized deidentification process.

This study proposed a 2-stage, end-to-end pipeline to automatically determine optic disc tilt from CFPs, as illustrated in Figure 1. The overall framework comprises 2 sequential stages: optic disc segmentation and tilt detection based on the ratio of the optic disc’s long and short axes.

In the segmentation process, a model was trained and validated on 3103 CFPs from the public Standardized Multi-Channel Dataset for Glaucoma (SMDG) [12], which contains optic disc annotations. The model was optimized to extract a binary segmentation mask representing the optic disc region. In the subsequent tilt detection step, the pretrained segmentation model was used to infer optic disc regions from the external validation dataset, which comprised 2448 CFPs acquired from SMC. For each predicted optic disc mask, the long diameter (LD) and short diameter (SD) were measured relative to the geometric center of the mask (see equations below) [13]. The tilt ratio was defined as the LD divided by the SD, and a cutoff of 1.3 was applied to classify the optic disc as tilted (LD:SD ratio ≥1.3) or nontilted (LD:SD ratio <1.3) [14,15].

The data used in this study were collected retrospectively from CFPs taken at the SMC Department of Ophthalmology between December 2007 and October 2021. A total of 1370 patients with 2448 CFPs were included in this study. The number of patients in each disease class was as follows: 514 healthy patients, 617 with glaucoma, 173 with pallor, and 66 with edema. The actual number of images in each disease class was as follows: 1215 normal, 894 glaucoma, 241 pallor, and 98 edema images. Images were excluded if their resolution was too low to allow clinical evaluation of the optic disc and related structures or if they showed severe artifacts, such as poor focus, strong light reflection, or motion blurring. The demographic and clinical characteristics of the study population are summarized in Table 1. The mean age was lowest in the normal group. The proportion of male patients ranged from 38% (n=66) to 55% (n=617), and the spherical equivalent refraction showed a tendency toward myopia in all groups.

To quantitatively evaluate segmentation performance, we used 3 metrics: the DSC (Eq. 3) [37], intersection over union (Eq. 4) [38], and pixel accuracy (Eq. 5). These metrics are defined as follows:

Where P and G denote the predicted segmentation mask and the ground-truth mask, respectively, and TP, TN, FP, and FN represent true-positive, true-negative, false-positive, and false-negative pixels, respectively. The 3 metrics range from 0 to 1, with values closer to 1 indicating higher similarity to the ground truth mask. These quantitative evaluations were performed using the SMDG, which provides ground-truth annotations. In contrast, the external validation dataset (ie, the SMC dataset) lacked annotation masks. To address this, we used the clinical acceptance rate as a human-based alternative for validating segmentation performance in the absence of ground-truth annotations. For the external validation dataset, segmentation masks were obtained by applying the pretrained segmentation model. The resulting masks were then evaluated independently by 2 ophthalmologists, and clinical acceptance rates were calculated separately for each reviewer to assess agreement with clinical judgment [39].

To identify the most suitable backbone for optic disc segmentation, we compared the performance of 4 representative 2D CNN–based architectures: U-Net [40], Attention U-Net [41], TransU-Net [42], and nnU-Net. All models were trained and evaluated under the same experimental conditions using the SMDG.

Table 5 summarizes the comparative performance of the models. Among the models, nnU-Net achieved the highest DSC and was selected as the backbone for the proposed pipeline. The results of 5-fold cross-validation for the selected nnU-Net model are summarized in Table 6. Performance was highly consistent across all 5 folds (DSC 0.956‐0.957), indicating stable generalizability without sensitivity to any particular data partition. Table 7 presents the per-dataset segmentation performance of both the single-fold model with the highest DSC among the 5 folds and the 5-fold homogeneous ensemble models across the 8 SMDG subdatasets. The single-fold model recorded a DSC >0.94, intersection over union >0.90, and pixel accuracy >0.997 across all datasets. The ensemble showed comparable but marginally lower performance than the single-fold model across most subdatasets. Given that accurate segmentation is a critical prerequisite for reliable tilt quantification, the single-fold model with the highest DSC among the 5 folds was selected for all subsequent tilt analyses.

Figure 2 presents visual examples of the segmentation results. The top row shows original CFPs, and the bottom row shows the corresponding predicted masks.

The performance of the optic disc tilt classification model was evaluated using the SMDG test set (n=620). Ground-truth labels were defined using the standard LD:SD ratio threshold of ≥ 1.3. At the clinically established threshold of an LD:SD ratio of ≥1.3, the model achieved an accuracy of 96.45% (598/620), a specificity of 98.27% (568/578), and a sensitivity of 71.43% (30/42). To characterize the distribution of misclassifications near the decision boundary, additional thresholds of ≥1.28 and ≥1.26 were evaluated (Table 8). The progressive recovery of sensitivity at lower thresholds indicates that misclassifications were concentrated in geometrically borderline cases. In these instances, the predicted LD:SD ratios were marginally lower (eg, 1.25-1.29) than the threshold, while the ground-truth ratios were slightly above 1.3. This pattern suggests that the misclassifications primarily reflect the limited representation of borderline tilt cases in the current dataset rather than a fundamental failure of the segmentation model itself, as the underlying segmentation maintained morphologic measurements that were very close to the ground truth. This dataset-driven limitation is expected to be mitigated in future work through the inclusion of additional borderline cases obtained through ongoing clinical data collection.

Two ophthalmologists independently reviewed each of the 2448 CFPs in the SMC dataset. To facilitate visual assessment, each CFP was displayed alongside its predicted segmentation mask, enabling the experts to determine clinical acceptability (1=“acceptable” and 0=“unacceptable”). The clinical acceptance rate was computed separately for each reviewer.

Both experts reported consistently high clinical acceptance rates across the normal (n=1209, 99.35% and n=1198, 98.61%), glaucoma (n=885, 99.11% and n=891, 99.67%), and pallor (n=240, 99.59% and n=239, 99.17%) groups, indicating reliable segmentation performance in these categories (Table 9). Acceptance rates were comparatively lower in the edema group (n=80, 82.63% and n=92, 93.88%), although most cases were still considered clinically acceptable by both reviewers. Across the entire dataset of 2448 images, the model achieved mean clinical acceptance rates of 98.61% (n=2414; expert 1) and 98.86% (n=2420; expert 2), with only 34 (1.39%) and 28 (1.14%) cases, respectively, assessed as clinically unacceptable.

The performance of the proposed optic disc tilt classification pipeline was evaluated on the external SMC dataset. Figure 3 illustrates representative tilt and nontilt classifications across the 4 disease types, demonstrating that the LD and SD were stably computed based on the segmentation results. A quantitative summary of the tilt classification distribution by disease type is provided in Table 10. Of the 2448 images, 186 cases (7.5%) were classified as tilt and 2262 cases (92.5%) as nontilt. Edema had the highest tilt ratio (14/98, 14.2%), followed by normal (118/1215, 9.7%) and pallor (19/241, 7.8%), with glaucoma showing the lowest incidence (35/894, 3.9%).

To further investigate the association between myopia severity and optic disc tilt, we performed a subgroup analysis of patients with high myopia (spherical equivalent refraction ≤−6.0 D), as summarized in Table 11. Interestingly, even within this high-myopia subgroup, the glaucoma group exhibited a lower tilt prevalence (15/86, 17.4%) than the normal group (75/225, 33.3%). This discrepancy can be largely attributed to the substantial age difference between the cohorts. The normal group, with a mean age of 11.1 years, consisted primarily of children and adolescents likely undergoing progressive axial elongation, a process strongly associated with the development of optic disc tilt. In contrast, the glaucoma group (mean age 58.6, SD 12.8 years) may have included patients whose myopia originated from different etiologic pathways not necessarily involving disc tilting. Furthermore, potential selection bias cannot be excluded; diagnosing glaucoma in highly myopic eyes with tilted discs is notoriously challenging because of altered optic nerve head morphology, which may have led to an underrepresentation of such complex cases in our retrospectively collected cohort.

Despite the overall high accuracy, we conducted a granular error analysis to acknowledge the study’s limitations. Recognition errors occurred in 1.4% (34/2448; expert 1) and 1.14% (28/2448; expert 2) of the cases in the external validation dataset, and unacceptable cases identified by expert 1 were categorized into 4 types (Table 12; Figure 4): swelling-induced, peripapillary atrophy−related, vessel boundary interference, and other unclassified errors. Notably, a reduced clinical acceptance rate was observed in the edema group, where pathological swelling distorted disc boundaries (Figure 4A), causing the model to misidentify the disc extent. This segmentation unreliability may naturally compromise the accuracy of downstream tilt estimation.

This study presents a robust AI-assisted pipeline for the automated detection and quantification of optic disc tilt, validated across diverse clinical conditions. By using an nnU-Net–based segmentation model trained on the SMDG dataset (3103 CFPs), we achieved outstanding performance that generalized effectively to the external SMC dataset (2448 CFPs), achieving mean clinical acceptance rates of 98.61% and 98.86% from 2 independent expert reviews. Our pipeline replaces subjective clinical judgments with an objective metric—the LD:SD ratio. Using a threshold of ≥1.3, the system identified 7.5% of images as tilted, with prevalence varying across the normal (9.7%), glaucoma (3.9%), pallor (7.8%), and edema (14.2%) groups. These findings affirm the model’s capacity to capture structural features across both healthy and diseased eyes, demonstrating robust boundary detection despite variations in image quality, race, and acquisition device.

The clinical implications of accurate optic disc tilt assessment are profound. Optic disc tilt occurs as a consequence of mechanical stress during axial elongation [14,43,44] and is correlated with the severity of myopic changes [4]. Tilted discs alter the 3D configuration of the optic nerve head, rendering traditional morphological criteria—such as the CDR and neuroretinal rim assessment—unreliable [6]. This structural distortion can mask genuine abnormalities or mimic pathological cupping, complicating the differentiation between physiological myopic changes and glaucoma [3]. Furthermore, tilt-induced alterations in retinal nerve fiber layer distribution can confound standard automated analyses [45,46]. By transitioning from manual, subjective assessments [47] to an automated, structure-based approach, our pipeline addresses these challenges. The integration of AI represents a paradigm shift in handling these geometric variations [48], overcoming the limitations of traditional algorithms that assume circular disc shapes [49]. The high agreement with expert judgment suggests that our system can reliably replicate human assessment [50] while reducing the interrater variability inherent in expert interpretations [51-53].

However, it is crucial to recognize that in clinical practice, the determination of optic disc tilt is of secondary importance in the presence of severe edema or papilledema. In such acute settings, the primary clinical objectives are the differential diagnosis of the underlying pathology and the preservation of vision rather than morphological profiling. Therefore, the inclusion of the edema group in this study was not intended to propose tilt assessment as a routine diagnostic step for these patients but rather to serve as a test of extreme conditions to evaluate the model’s robustness and boundary detection limits under challenging morphological conditions. Similarly, peripapillary atrophy regions mimicking disc intensity (Figure 4B) occasionally led to false boundary extensions. These findings highlight the inherent challenge of applying a model trained exclusively on public datasets to severe pathological cases. Future iterations could mitigate these errors through training with more diverse, disease-specific annotations, contributing to ongoing research on tilt classification [54,55]. In conclusion, this AI-assisted tilt detection system offers significant potential for enhancing clinical practice. It facilitates efficient, large-scale screening for myopia-related abnormalities and enables longitudinal monitoring of tilt progression. The objective quantification provided by this tool is particularly valuable for telemedicine applications, extending expert-level assessment to underserved regions [56]. Furthermore, tilt-aware diagnostics could reduce health care costs by preventing unnecessary treatments driven by false-positive findings [57]. Future research should focus on integrating multimodal imaging to further elucidate the pathophysiology of tilt-related complications. By providing a robust, reproducible solution to a complex diagnostic problem, this study represents a meaningful step toward precision medicine in ophthalmology.

This study validated an AI-based pipeline for the automated detection and measurement of optic disc tilt on CFPs. By replacing manual segmentation and subjective clinical assessments, the pipeline provides objective and reliable tilt measurements, creating new possibilities for screening and diagnosing myopia-related eye conditions. The pipeline performed consistently across various clinical conditions (ie, normal, glaucoma, pallor, and edema) and effectively identified optic disc tilt in populations with a high prevalence of myopia. This system enhances the standardization and scalability of AI-driven diagnostics and shows potential for integration into telemedicine and large-scale screening programs. Future improvements, such as addressing challenges in edema cases and incorporating multimodal imaging, could further improve diagnostic accuracy. This work lays a strong foundation for improving the quality and accessibility of care for myopia-related eye diseases.