Translational Vision Science & Technology Cover Image for Volume 14, Issue 6
June 2025
Volume 14, Issue 6
Open Access
Artificial Intelligence  |   June 2025
Comparison of AI-Automated and Manual Subfoveal Choroidal Thickness Measurements in an Elderly Population Using Optical Coherence Tomography
Author Affiliations & Notes
  • Wen-Da Zhou
    Beijing Tongren Eye Center, Beijing key Laboratory of Intraocular Tumor Diagnosis and Treatment, Beijing Ophthalmology & Visual Sciences Key Lab, Medical Artificial Intelligence Research and Verification Key Laboratory of the Ministry of Industry and Information Technology, Beijing Key Laboratory of Intelligent Diagnosis, Treatment and Prevention of Blinding Eye Diseases, Beijing Tongren Hospital, Capital Medical University, Beijing, China
  • Han-Qing Zhao
    Beijing Tongren Eye Center, Beijing key Laboratory of Intraocular Tumor Diagnosis and Treatment, Beijing Ophthalmology & Visual Sciences Key Lab, Medical Artificial Intelligence Research and Verification Key Laboratory of the Ministry of Industry and Information Technology, Beijing Key Laboratory of Intelligent Diagnosis, Treatment and Prevention of Blinding Eye Diseases, Beijing Tongren Hospital, Capital Medical University, Beijing, China
  • Jia-Qi Geng
    EVision Technology (Beijing) Co. LTD, Beijing, China
  • Yu-Hang Yang
    Beijing Tongren Eye Center, Beijing key Laboratory of Intraocular Tumor Diagnosis and Treatment, Beijing Ophthalmology & Visual Sciences Key Lab, Medical Artificial Intelligence Research and Verification Key Laboratory of the Ministry of Industry and Information Technology, Beijing Key Laboratory of Intelligent Diagnosis, Treatment and Prevention of Blinding Eye Diseases, Beijing Tongren Hospital, Capital Medical University, Beijing, China
  • Li Dong
    Beijing Tongren Eye Center, Beijing key Laboratory of Intraocular Tumor Diagnosis and Treatment, Beijing Ophthalmology & Visual Sciences Key Lab, Medical Artificial Intelligence Research and Verification Key Laboratory of the Ministry of Industry and Information Technology, Beijing Key Laboratory of Intelligent Diagnosis, Treatment and Prevention of Blinding Eye Diseases, Beijing Tongren Hospital, Capital Medical University, Beijing, China
  • Rui-heng Zhang
    Beijing Tongren Eye Center, Beijing key Laboratory of Intraocular Tumor Diagnosis and Treatment, Beijing Ophthalmology & Visual Sciences Key Lab, Medical Artificial Intelligence Research and Verification Key Laboratory of the Ministry of Industry and Information Technology, Beijing Key Laboratory of Intelligent Diagnosis, Treatment and Prevention of Blinding Eye Diseases, Beijing Tongren Hospital, Capital Medical University, Beijing, China
  • Wen-Bin Wei
    Beijing Tongren Eye Center, Beijing key Laboratory of Intraocular Tumor Diagnosis and Treatment, Beijing Ophthalmology & Visual Sciences Key Lab, Medical Artificial Intelligence Research and Verification Key Laboratory of the Ministry of Industry and Information Technology, Beijing Key Laboratory of Intelligent Diagnosis, Treatment and Prevention of Blinding Eye Diseases, Beijing Tongren Hospital, Capital Medical University, Beijing, China
  • Lei Shao
    Beijing Tongren Eye Center, Beijing key Laboratory of Intraocular Tumor Diagnosis and Treatment, Beijing Ophthalmology & Visual Sciences Key Lab, Medical Artificial Intelligence Research and Verification Key Laboratory of the Ministry of Industry and Information Technology, Beijing Key Laboratory of Intelligent Diagnosis, Treatment and Prevention of Blinding Eye Diseases, Beijing Tongren Hospital, Capital Medical University, Beijing, China
  • Correspondence: Wen-Bin Wei, Beijing Tongren Eye Center, Beijing Key Laboratory of Intraocular Tumor Diagnosis and Treatment, Beijing Ophthalmology & Visual Sciences Key Lab, Medical Artificial Intelligence Research and Verification Key Laboratory of the Ministry of Industry and Information Technology, Beijing Tongren Hospital, Capital Medical University, Dongjiaominxiang Street, No 1, Dongcheng District, Beijing 100730, China. e-mail: [email protected] 
  • Lei Shao, Beijing Tongren Eye Center, Beijing Key Laboratory of Intraocular Tumor Diagnosis and Treatment, Beijing Ophthalmology & Visual Sciences Key Lab, Medical Artificial Intelligence Research and Verification Key Laboratory of the Ministry of Industry and Information Technology, Beijing Tongren Hospital, Capital Medical University, Dongjiaominxiang Street, No 1, Dongcheng District, Beijing 100730, China. e-mail: [email protected] 
  • Footnotes
     WDZ and HQZ contributed equally to this study and share first authorship.
Translational Vision Science & Technology June 2025, Vol.14, 9. doi:https://doi.org/10.1167/tvst.14.6.9
  • Views
  • PDF
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Wen-Da Zhou, Han-Qing Zhao, Jia-Qi Geng, Yu-Hang Yang, Li Dong, Rui-heng Zhang, Wen-Bin Wei, Lei Shao; Comparison of AI-Automated and Manual Subfoveal Choroidal Thickness Measurements in an Elderly Population Using Optical Coherence Tomography. Trans. Vis. Sci. Tech. 2025;14(6):9. https://doi.org/10.1167/tvst.14.6.9.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose: To evaluate the agreement and correlation between manual and automated measurements of subfoveal choroidal thickness (SFCT) using enhanced depth imaging spectral-domain optical coherence tomography in an elderly population and to investigate the factors influencing measurement discrepancies.

Methods: Based on the Beijing Eye Study, SFCT was measured manually using Heidelberg Eye Explorer software and automatically via a TransUNet-based deep learning model. Agreement between manual and automated SFCT measurements was assessed using Bland-Altman plots, intraclass correlation coefficients (ICC), and Pearson correlation coefficients.

Results: Among 2896 participants, automated and manual measurements of SFCT demonstrated strong correlation (ICC = 0.971; 95% confidence interval [CI], 0.969–0.973; Pearson = 0.974, P < 0.001). Subgroup analyses showed similarly high correlation across participants aged ≥60 years (ICC = 0.954, Pearson = 0.974), aged <60 years (ICC = 0.971; Pearson = 0.953), with axial length ≥23 mm (ICC = 0.969; Pearson = 0.974), and axial length <23 mm (ICC = 0.959; Pearson = 0.963). Participants with SFCT <300 µm showed higher consistency (ICC = 0.942; Pearson = 0.944) compared to those with SFCT ≥300 µm (ICC = 0.867; Pearson = 0.868). Significant fixed and proportional biases were observed in all subgroups (P < 0.001), with manual measurements consistently lower than automated values.

Conclusions: Despite the presence of systematic biases, automated SFCT measurements showed excellent consistency and strong correlation with manual measurements across a large elderly population. These findings support the potential utility of AI-assisted SFCT measurement in clinical settings.

Translational Relevance: This study validates AI-based SFCT measurement in a large elderly cohort, enhancing diagnostic accuracy and bridging research with practice.

Introduction
The choroid, a vascular layer located between the sclera and retina, receives the majority of ocular blood flow and plays a crucial role in the pathogenesis of various posterior segment eye diseases, including age-related macular degeneration, polypoidal choroidal vasculopathy, central serous chorioretinopathy, and myopic retinopathy.15 Accurate and reliable assessment of subfoveal choroidal thickness (SFCT) provides valuable insights into structural changes within the choroid, which are associated with a variety of ocular diseases, making SFCT measurement a crucial parameter for diagnosing and monitoring the progression of these conditions in clinical settings.68 Optical coherence tomography (OCT) has emerged as a preferred tool for noninvasive imaging of the choroidal structure, offering high-resolution cross-sectional images of the retina and underlying layers.9,10 However, although OCT provides detailed anatomical information, manual measurement of SFCT remains time-consuming, requires substantial expertise, and is prone to operator variability, introducing measurement bias and potentially affecting the consistency of clinical follow-ups and research findings. 
In recent years, artificial intelligence (AI) has gained significant attention in ophthalmology, particularly for automating the analysis of OCT images. AI-based systems, especially those using deep learning models, have shown promise in delivering faster and more consistent measurements while minimizing human intervention.1116 These technologies hold the potential to revolutionize SFCT measurement, improving both precision and efficiency. Previous studies have evaluated the consistency between manual and automated SFCT measurements; however, most have been limited by small sample sizes, restricting the generalizability of their findings. A more recent study, which assessed the agreement between manual and automated SFCT measurements using swept-source OCT (SS-OCT) in a larger cohort of adult myopic individuals, demonstrated good consistency.17 Although this study was an important step in validating automated measurements, it was restricted to middle-aged and young myopic individuals, a group that may not fully represent the broader population. 
Thus the present study aims to compare automated and manual SFCT measurements using enhanced depth imaging spectral-domain OCT images in a large cohort of elderly individuals, addressing the need for accurate and reproducible methods in clinical practice. To our knowledge, this is the first large-scale study focused specifically on the elderly population, considering a wide range of systemic and ocular factors that may influence SFCT. By examining these variables, we aim to evaluate the clinical applicability of AI-driven SFCT measurements, providing valuable insights into its potential as a reliable tool for both clinical and research purposes in geriatric ophthalmology. 
Method
We conducted a cross-sectional study using subjects recruited from the Beijing Eye Study 2011. The Beijing Eye Study 2011 is a population-based cross-sectional study conducted in northern China in the year 2011. It included 3468 individuals (1963 women, 56.6%) with a mean age of 64.6 ± 9.8 years (range 50–93 years).18,19 
For this study, we included subjects with available OCT assessments meeting all of the following quality criteria: The OCT image needed to be centered scans with adequate fixation and clear delineation of all retinal layers without boundary disruption. Subjects were excluded if they had any ocular media opacity affecting image quality (such as corneal opacities, clinically significant cataracts, or vitreous opacities); pathologic retinal conditions or any retinal lesion causing retinal layer disruption (such as retinal atrophy, macular edema, epiretinal membranes, macular holes, or retinal detachments); or any iatrogenic factors (postsurgical changes or laser scars). High myopia, especially myopic macular degeneration, is another critical factor affecting retinal and choroidal architecture in OCT imaging. To address this, all tractional and neovascular degeneration cases were excluded. For the atrophic degeneration, we excluded eyes exhibiting Bruch's membrane defects (i.e., patchy atrophy and macular atrophy), because these lesions significantly disrupt retinal layer integrity. In contrast, fundus tessellation and diffuse atrophy—which primarily cause choroidal thinning without obscuring the boundaries of retinal layers on OCT—were retained in the study cohort. The number of high axial myopia (AL > 26 mm) individuals included in this study was 16, with an axial length of 26.5 ± 0.64 mm (range 26.01–28.68 mm). The study protocol was approved by the Medical Ethics Committee of Beijing Tongren Hospital. Written informed consent was obtained from all participants, and the study adhered to the tenets of the Declaration of Helsinki. 
All study participants were interviewed using a standardized questionnaire regarding their family status, educational level, income, quality of life, psychological depression, physical activity levels, and prevalent systemic diseases, such as arterial hypertension and diabetes mellitus. The ophthalmic examination included measurements of presenting visual acuity, uncorrected visual acuity, and best-corrected visual acuity, as well as tonometry and slit-lamp examination of the anterior segment. Axial length and other biometric parameters were measured using optical low-coherence reflectometry with the Lenstar 900 Optical Biometer (Haag-Streit, Koeniz, Switzerland). Digital photography was used to document the cornea, lens, macula, and optic disc. 
SFCT was measured using spectral-domain optical coherence tomography (SD-OCT; Spectralis, Wavelength: 870 nm; Heidelberg Engineering Co., Heidelberg, Germany) with the enhanced depth imaging modality after pupil dilation. Seven sections, each comprising 100 averaged scans, were obtained within a 58° to 308° rectangular area centered on the fovea. For this study, only OCT B-scan images that passed through the foveal center were selected for both manual annotation and automated measurement. The horizontal section passing through the foveal center was selected for further analysis. SFCT was defined as the vertical distance from the hyperreflective line of Bruch's membrane to the hyperreflective line of the choroid–scleral interface (CSI). Manual measurements were performed using the Heidelberg Eye Explorer software (version 5.3.3.0, Heidelberg Engineering Co.). All images were captured by a single technician and independently assessed in a masked manner by two ophthalmologists. The final result was obtained by averaging their measurements. The manual technique has been previously described and demonstrated excellent agreement and reproducibility.20 
Automated SFCT measurement based on the artificial intelligence analysis platform EVisionAI, OCT image layer segmentation, macular fovea localization, and SFCT calculation were carried out. We used a TransUNet network model, which follows a U-shaped encoder-decoder architecture consisting of an encoder, a Transformer module, and a decoder.21 In the field of medical imaging, the resolution of OCT images often reaches a relatively high pixel level. When CNN networks process large-sized images and conduct deep convolutional neural network analysis, they still face the inherent limitations of convolution operations and perform poorly in the interaction of global semantic information. TransUnet combines the advantages of Transformer and U-Net, using Transformer as the encoder to capture global context information through its self-attention mechanism, whereas U-Net is responsible for precise feature extraction and the restoration of image details. This combination not only enhances the model's ability to understand global information but also maintains its high sensitivity to local details. In terms of the model's generalization ability, TransUnet performs better, especially for OCT images with lesions. The encoder first downsamples the input image using convolution operations, embeds image blocks from the downsampled feature map, adds positional encoding, and then inputs the resulting one-dimensional vector into the Transformer structure, which consists of 12 layers. The Transformer module processes Layer Normalization, Multi-head Self-Attention, Multi-Layer Perceptron, and residual connection structures for the embedded image block vectors. In the decoder, transposed convolution is used for upsampling to recover image pixels, while cascading feature maps of the same resolution as the downsampled ones from the encoder. To detect the macular fovea, we labeled the detection box and used the SSD (Single Shot MultiBox Detector), a classic single-stage, anchor-based target detection model.22 The SSD model uses VGG16 as its base network, with modifications to adapt it for target detection.23 These include converting the fully connected layers FC6 and FC7 into convolutional layers (Conv6 and Conv7), removing all Dropout layers and the FC8 layer, and adding new convolutional layers to generate additional feature maps. The SSD model also adds multiscale feature blocks to extract features at different scales and uses class and bounding box prediction layers to calculate the probability of each anchor box belonging to a particular class and adjust the bounding boxes to more accurately locate the target. Non-maximum suppression is applied to remove overlapping bounding boxes and retain the one with the highest confidence as the final result. Because of tilt during image capture, geometric correction was performed by first calculating the tilt angle, then extracting the centerline of the fovea detection box and determining the intersection of the uppermost retinal layer with this centerline as the macular fovea. Finally, the thickness of the retina and choroid below the fovea was calculated. A total of 300 OCT images were used as a validation set to evaluate the performance of the model in measuring SFCT. The accuracy, sensitivity, and specificity of the automated measurements, compared with manual measurements, were 0.958, 0.937, and 0.960, respectively. Figure 1 shows the results of automated segmentation performed by the TransUnet model. 
Figure 1.
 
Automated choroidal segmentation. (A) OCT B-scan images that passed through the foveal center; (B) automated choroidal segmentation generated by TransUnet model.
Figure 1.
 
Automated choroidal segmentation. (A) OCT B-scan images that passed through the foveal center; (B) automated choroidal segmentation generated by TransUnet model.
All statistical analyses were conducted using SPSS (version 25; IBM, Chicago, IL, USA). Descriptive statistics were calculated for participants who met the inclusion criteria. To compare subject characteristics between different age groups (<60 years vs. ≥60 years) and axial length groups (<23 mm vs. ≥23 mm), independent t-tests were used for continuous variables, whereas χ2 or Fisher's exact tests were applied to categorical variables. The agreement between automated and manual measurements of SFCT was assessed using Bland-Altman plots, with the difference between manual and automated measurements (manual minus automated) plotted against the mean of both measurements. The proportion of outliers was calculated by dividing the number of data points outside the 95% limits of agreement (LOA) by the total number of measurements. Pearson's correlation coefficient was used to evaluate the strength of the relationship between automated and manual choroidal thickness measurements. The intraclass correlation coefficient (ICC) was calculated using a two-way mixed-effects model with absolute agreement to assess the degree of consistency between the two measurement methods. Additionally, a one-sample t-test and linear regression model were performed to evaluate the presence of fixed and proportional bias, respectively. Subgroup analyses for correlation and agreement were performed based on age (<60 years vs. ≥60 years), axial length (<23.0 mm vs. ≥23.0 mm), and choroidal thickness (<300 µm vs. ≥300 µm). 
Results
Among the 3234 participants with available OCT scans, 2896 individuals (89.5%) were included in the final analysis. Of these, 1323 (45.7%) were men, and the mean age was 63.9 ± 6.9 years (range 43–93 years). OCT imaging and SFCT measurements were obtained from the right eye of each participant. Table 1 shows the baseline characteristics between included and excluded participants, with subgroup analysis of included participants stratified by age and axial length. The included participants had a mean axial length of 23.2 ± 0.9 mm, a spherical equivalent (SE) of 0.00 ± 1.6 diopters, an intraocular pressure of 14.5 ± 2.8 mm Hg, and a best-corrected visual acuity (BCVA) of 0.06 ± 0.17 LogMAR. Their mean height, weight, systolic blood pressure, diastolic blood pressure, and heart rate were 162.3 ± 8.7 cm, 67.3 ± 12.0 kg, 129.3 ± 20.3 mm Hg, 69.5 ± 12.2 mm Hg, and 72.5 ± 12.4 beats/min, respectively. The prevalence of hypertension and diabetes in the included group was 45.7% and 12.2%, respectively. In comparison, participants who were excluded from the analysis were significantly older (65.6 ± 10.7 vs. 63.9 ± 6.9 years, P = 0.001), had longer axial length (23.7 ± 1.7 mm vs. 23.2 ± 0.9 mm, P < 0.001), more myopic refractive status (−1.09 ± 3.6 D vs. 0.00 ± 1.6 D, P < 0.001), and poorer BCVA (0.18 ± 0.39 vs. 0.06 ± 0.17 LogMAR, P < 0.001). In addition, the prevalence of diabetes was significantly higher among excluded participants (21.6% vs. 12.2%, P < 0.001). However, no statistically significant differences were observed in intraocular pressure, body weight, diastolic blood pressure, heart rate, or prevalence of hypertension between the included and excluded groups (all P > 0.05). Compared with younger participants, those aged 60 years or older showed significantly poorer BCVA (0.10 ± 0.19 vs. 0.003 ± 0.10 LogMAR), longer axial length (23.3 ± 1.0 mm vs. 23.1 ± 0.94 mm), lower intraocular pressure (14.2 ± 2.8 mm Hg vs. 14.9 ± 2.7 mm Hg), shorter height (161.4 ± 1.8 cm vs. 163.7 ± 8.3 cm), lower body weight (65.7 ± 12.0 kg vs. 69.6 ± 11.6 kg), higher systolic blood pressure (132.1 ± 20.4 mm Hg vs. 124.9 ± 19.5 mm Hg), lower diastolic blood pressure (68.3 ± 11.9 mm Hg vs. 71.4 ± 12.5 mm Hg), and a higher prevalence of both hypertension (53.5% vs. 34.0%) and diabetes (15.4% vs. 7.4%) (all P < 0.001). Participants with an axial length of 23 mm or greater were more likely to be male and had significantly older age (64.5 ± 9.8 years vs. 62.7 ± 9.3 years), a more myopic spherical equivalent (−0.40 ± 1.74 diopters vs. 0.53 ± 1.26 diopters), greater height (164.5 ± 8.7 cm vs. 159.5 ± 7.5 cm), higher body weight (68.9 ± 11.9 kg vs. 65.5 ± 11.3 kg), and lower systolic blood pressure (128.2 ± 20.0 mm Hg vs. 130.8 ± 21.1 mm Hg) (all P < 0.01). 
Table 1.
 
Comparison of Baseline Characteristics Between Included and Excluded Participants, With Subgroup Analysis of Included Participants Stratified by Age and Axial Length
Table 1.
 
Comparison of Baseline Characteristics Between Included and Excluded Participants, With Subgroup Analysis of Included Participants Stratified by Age and Axial Length
The agreement and correlation between manual and automated measurements of SFCT are summarized in Table 2. Among all 2896 participants, a strong correlation was observed between the two methods, with an ICC of 0.971 (95% confidence interval [CI], 0.969–0.973; P < 0.001) and a Pearson correlation coefficient of 0.976 (P < 0.001) (Fig. 2). The 95% LOA ranged from −84.77 µm to 23.91 µm, with 4.8% of measurements falling outside these limits (Fig. 3). A significant fixed bias was observed, with manual measurements being on average 30.42 µm lower than automated measurements (95% CI, −31.43 to −29.44, P < 0.001). A significant proportional bias was also present (B = −25.75; 95% CI, −28.60 to −22.90; P < 0.01). When stratified by age, participants aged ≥60 years had an ICC of 0.954 (95% CI, 0.946–0.963, P < 0.001) and a Pearson correlation coefficient of 0.974 (P < 0.001), while those aged <60 years demonstrated an even higher ICC of 0.971 (95% CI, 0.967–0.972; P < 0.001) and a Pearson correlation coefficient of 0.953 (P < 0.001), indicating strong consistency between the two methods in both age groups. Significant fixed and proportional biases were observed in both groups (all P < 0.001), with slightly greater fixed biases in the younger participants (−32.12 µm vs. −29.35 µm). When stratified by axial length, participants with axial length ≥23 mm had an ICC of 0.969 (95% CI, 0.966–0.972; P < 0.001) and a Pearson correlation coefficient of 0.974 (P < 0.001), whereas those with axial length <23 mm also showed strong consistency with an ICC of 0.959 (95% CI, 0.954–0.963; P < 0.001) and a Pearson correlation coefficient of 0.963 (P < 0.001). In both groups, significant fixed and proportional biases were detected (all P < 0.001). The group with axial length <23 mm demonstrated a slightly larger fixed bias (−32.88 µm vs. −28.59 µm). When stratified by SFCT level, participants with SFCT ≥300 µm showed relatively lower agreement, with an ICC of 0.867 (95% CI, 0.852–0.882; P < 0.001) and a Pearson correlation coefficient of 0.868 (P < 0.001), although the consistency remained strong. In contrast, participants with SFCT <300 µm demonstrated stronger correlation, with an ICC of 0.942 (95% CI, 0.937–0.947; P < 0.001) and a Pearson correlation coefficient of 0.944 (P < 0.001), along with narrower limits of agreement (−73.91 µm to 13.11 µm) and a lower proportion of outliers (4.6%). Significant fixed and proportional biases were observed in both SFCT subgroups (all P < 0.001). 
Table 2.
 
Agreement and Correlation Between Manual and Automated Measurements of SFCT Stratified by Age, Axial Length, and SFCT
Table 2.
 
Agreement and Correlation Between Manual and Automated Measurements of SFCT Stratified by Age, Axial Length, and SFCT
Figure 2.
 
Scatter plot showing the correlation between the automated and manual subfoveal choroidal thickness measurements performed on optical coherence tomograms.
Figure 2.
 
Scatter plot showing the correlation between the automated and manual subfoveal choroidal thickness measurements performed on optical coherence tomograms.
Figure 3.
 
Bland–Altman plot demonstrating the agreement between automated and manual subfoveal choroidal thickness measurements.
Figure 3.
 
Bland–Altman plot demonstrating the agreement between automated and manual subfoveal choroidal thickness measurements.
Discussion
In this study, we aimed to evaluate the agreement and correlation between automated and manual measurements of SFCT in a large cohort of elderly individuals. Our results showed an ICC of 0.971 and a Pearson correlation coefficient (r) of 0.974, indicating a high level of consistency between the two methods across the entire study population. Previous studies have suggested that axial length can influence optical magnification and the accuracy of OCT-derived ocular parameters due to transverse magnification effects.24,25 To evaluate whether ocular magnification related to axial length influenced the agreement between manual and automated measurements, we conducted a subgroup analysis by stratifying participants into two axial length groups (≥23 mm and <23 mm). The results revealed that although fixed and proportional biases were present in both subgroups, the overall Pearson correlation coefficient and ICC remained consistent, indicating that axial length had a minimal effect on the measurement discrepancies. This suggests that the observed differences between the two methods were more likely attributable to inherent limitations in measurement techniques rather than ocular magnification alone. Moreover, precise foveal center localization was a key component of the AI-model, which could help to mitigate the impact of lateral ocular magnification associated with axial length and ensured reliable targeting of the subfoveal region for consistent thickness measurement. Similarly, age-related changes in choroidal thickness have been well-documented in the literature.26 Thus we also stratified our analysis by age to examine its effect on the agreement between the two methods. Again, we found minimal differences in the correlation and agreement indices, indicating that both manual and automated methods were equally influenced by age-related changes. Furthermore, we stratified our analysis by choroidal thickness and observed weaker correlation and agreement indices in subjects with thicker choroids (≥300 µm). Although the correlation remained high, with ICC and Pearson correlation coefficients of 0.867 and 0.868, respectively, the magnitude of correlation was notably lower in this group. This finding is consistent with previous studies conducted in myopic adult cohorts, which also reported reduced agreement in individuals with thicker choroids.17 The larger fixed and proportional bias observed in eyes with thicker choroids suggests that the difference between manual and automated measurements was more pronounced in this subgroup. 
It is particularly noteworthy that this study identified significant fixed and proportional biases between manual and automated measurements of SFCT. Specifically, manual measurements tended to yield lower SFCT values compared to those obtained by the AI, and the discrepancy increased with greater choroidal thickness. We believe this phenomenon is primarily attributable to the inherent limitations of manual measurement techniques and the technical constraints of the OCT system employed in this study. Specifically, we used an enhanced depth imaging spectral-domain OCT, which represented state-of-the-art technology at the time. However, its limited tissue penetration compared to more recent SS-OCT systems may have hindered the accurate visualization of deeper choroidal boundaries.27 It is important to note that manual SFCT measurement is not a gold standard and is inherently subject to inter- and intra-observer variability. In our study, manual measurements were performed using the built-in caliper function of the Heidelberg software, which does not provide automated layer segmentation. As a result, observers were required to manually identify the CSI. However, because of the limited penetration capability of SD-OCT, the CSI beneath the fovea often appeared indistinct. In such cases, observers had to extrapolate the CSI line from peripheral regions—where the choroid is thinner and the interface more clearly visualized—toward the foveal center by simulating the curvature of the posterior scleral surface. These limitations are likely to have a greater impact in eyes with thicker choroids, where the CSI is more difficult to define, thereby contributing to the proportional bias observed between manual and automated measurements. Our findings are consistent with those of Lim et al.,17 who used a swept-source OCT (Triton DRI-OCT) to compare manual and AI-derived SFCT in a predominantly young, highly myopic population (mean age, 34.1 ± 10.4 years; mean axial length, 26.4 ± 1.4 mm). Their study also reported high correlation but significant fixed and proportional biases between the two methods. Notably, no significant proportional bias was observed in subjects with SFCT <300 µm, further suggesting that choroidal thickness may be a key factor influencing systematic measurement bias. In contrast to their population, the elderly participants in our study exhibited slightly thicker choroids (mean SFCT, 264.2 ± 106.4 µm vs. 239.3 ± 84.3 µm), and the lower tissue penetration of the SD-OCT system used may have compounded measurement challenges in thicker choroids. 
The high ICC and correlation coefficients suggest a strong consistency and correlation in measurement trends between the two methods; however, the presence of significant fixed and proportional biases indicates systematic disagreement, suggesting that the two methods exhibit strong relative agreement but limited absolute agreement. Although the significant fixed and proportional biases suggest that the two measurement methods may not be directly interchangeable, it is worth emphasizing that choroidal thickness is a highly dynamic parameter, subject to fluctuations influenced by factors such as systemic blood pressure, circadian rhythm, dietary intake, and ambient light exposure.28 Given this inherent variability, the precise absolute value of choroidal thickness may be less critical in clinical settings. Instead, longitudinal intraindividual comparisons are more meaningful for disease monitoring and management.29 Therefore, in practical applications, using a consistent OCT device and a uniform measurement approach is generally sufficient. In this light, despite the presence of systematic bias, the strong correlation between manual and automated measurements underscores the potential of AI-based methods as a reliable tool for clinical follow-up. 
In clinical practice, the ability to longitudinally monitor changes in choroidal thickness is crucial for the long-term management of chronic retinal conditions in the elderly, such as age-related macular degeneration, diabetic retinopathy, and hypertensive retinopathy—all of which have potential association with changes in choroidal structure.4,2932 However, current clinical assessments of choroidal thickness largely rely on qualitative evaluation by ophthalmologists, because precise manual measurement is time-consuming, costly, and requires standardized training, placing a significant burden on clinical workflows. Although some advanced OCT systems are equipped with automated choroidal segmentation and quantification features, their availability remains limited, particularly in primary care settings. Most community hospitals in China still rely on older SD-OCT systems. Given the growing demand for eye health services in China's aging population, our findings suggest that even with SD-OCT technology, AI-based measurement systems have the potential to effectively substitute manual SFCT measurements. This could substantially alleviate clinical workloads, especially in resource-limited areas where access to trained ophthalmologists is scarce, ultimately improving eye health management in the elderly.33 
The limitations of this study should still be acknowledged. Although our study included a relatively large sample size, the proportion of participants with high myopia (axial length > 26 mm), including those with pathological myopia, was low. As a result, the generalizability of our model to these populations remains uncertain and warrants further validation in future studies using relevant datasets. 
In conclusion, our study demonstrates that AI-assisted measurement of SFCT using SD-OCT images is strongly correlated with manual measurement, despite the presence of significant systematic bias. Given the dynamic nature of choroidal thickness and the demands of clinical practice, AI-based measurements provide a promising alternative to manual assessments, with potential to support disease monitoring and long-term management of retinal conditions in older adults. Further validation in diverse populations and with newer imaging technologies, such as SS-OCT, may help to improve the absolute agreement and expand the clinical applicability of AI-based choroidal measurements. 
Acknowledgments
Supported by the National Natural Science Foundation of China (82220108017, 82141128, 82401283); The Capital Health Research and Development of Special (2024-1-2052); Science & Technology Project of Beijing Municipal Science & Technology Commission (Z201100005520045); Sanming Project of Medicine in Shenzhen (No. SZSM202311018); Scientific Research Common Program of Beijing Municipal Commission of Education (No. KM202410025011); The priming scientific research foundation for the junior researcher in Beijing Tongren Hospital, Capital Medical University (No. 2023-YJJ-ZZL-003). 
Author Contributions: Concept and design: Wen-Da Zhou, Han-qing Zhao, Jia-Qi Geng, Lei Shao, Wen-Bin Wei. Acquisition, analysis, or interpretation of data: Wen-Da Zhou, Yu-hang Yang, Han-qing Zhao, Li Dong, Ruiheng Zhang. Critical revision of the manuscript for important intellectual content: All authors. 
Data Availability and Materials: The data that support the findings of this study are not publicly available due to their containing information that could compromise the privacy of research participants but are available from the corresponding author upon reasonable request. 
Disclosure: W.-D. Zhou, None; H.-Q. Zhao, None; J.-Q. Geng, EVision Technology LTD (E); Y.-H. Yang, None; L. Dong, None; R. Zhang, None; W.-B. Wei, None; L. Shao, None 
References
Fekrazad S, Shahrabi Farahani M, Salehi MA, Hassanzadeh G, Arevalo JF. Choroidal thickness in eyes of rheumatoid arthritis patients measured using optical coherence tomography: a systematic review and meta-analysis. Surv Ophthalmol. 2024; 69: 435–440. [CrossRef] [PubMed]
Tan CS, Cheong KX. Macular choroidal thicknesses in healthy adults–relationship with ocular and demographic factors. Invest Ophthalmol Vis Sci. 2014; 55: 6452–6458. [CrossRef] [PubMed]
Chen Y, Yuan Y, Zhang S, et al. Retinal nerve fiber layer thinning as a novel fingerprint for cardiovascular events: results from the prospective cohorts in UK and China. BMC Med. 2023; 21(1): 24. [CrossRef] [PubMed]
Shao L, Zhou LX, Xu L, Wei WB. The relationship between subfoveal choroidal thickness and hypertensive retinopathy. Sci Rep. 2021; 11(1): 5460. [CrossRef] [PubMed]
Gouravani M, Salehi MA, Mohammadi S, Arevalo JF. Choroidal thickness in eyes of migraine patients measured using spectral domain-optical coherence tomography: A meta-analysis. Surv Ophthalmol. 2023; 68: 67–77. [CrossRef] [PubMed]
Kazantzis D, Machairoudia G, Theodossiadis P, Chatziralli I. Subfoveal choroidal thickness changes in patients with pseudoexfoliation syndrome (PEX) compared to healthy controls: a systematic review and meta-analysis. Photodiagn Photodyn Ther. 2024; 47: 104095. [CrossRef]
Biyik KZ, Tideman JWL, Polling JR, et al. Subfoveal choroidal thickness at age 9 years in relation to clinical and perinatal characteristics in the population-based Generation R Study. Acta Ophthalmol. 2020; 98: 172–176. [CrossRef] [PubMed]
Ala-Kauhaluoma M, Koskinen SM, Silvennoinen H, et al. Subfoveal choroidal thickness in ipsi- and contralateral eyes of patients with carotid stenosis before and after carotid endarterectomy: a prospective study. Acta Ophthalmol. 2021; 99: 545–552. [CrossRef] [PubMed]
Zhang L, Buitendijk GH, Lee K, et al. Validity of automated choroidal segmentation in SS-OCT and SD-OCT. Invest Ophthalmol Vis Sci. 2015; 56: 3202–3211. [CrossRef] [PubMed]
Adhi M, Liu JJ, Qavi AH, et al. Choroidal analysis in healthy eyes using swept-source optical coherence tomography compared to spectral domain optical coherence tomography. Am J Ophthalmol. 2014; 157: 1272–1281.e1271. [CrossRef] [PubMed]
Song D, Ni Y, Zhou Y, et al. Evaluation of choroid vascular layer thickness in wet age-related macular degeneration using artificial intelligence. Photodiagn Photodyn Ther. 2024; 47: 104218. [CrossRef]
Lin CY, Huang YL, Hsia WP, Wang Y, Chang CJ. Correlation of choroidal thickness with age in healthy subjects: automatic detection and segmentation using a deep learning model. Int Ophthalmol. 2022; 42: 3061–3070. [CrossRef] [PubMed]
Li Y, Wong D, Sreng S, et al. Effect of childhood atropine treatment on adult choroidal thickness using sequential deep learning-enabled segmentation. Asia Pac J Ophthalmol (Phila). 2024; 13(5): 100107. [CrossRef] [PubMed]
Li J, Zhu L, Zhu R, et al. Automated analysis of choroidal sublayer morphologic features in myopic children using EDI-OCT by deep learning. Transl Vis Sci Technol. 2021; 10(13): 12. [CrossRef] [PubMed]
Li X, Liu G, Ni Y, et al. Analysis of macula structural changes in Moyamoya disease using AI-assisted OCT. Photodiagn Photodyn Ther. 2024; 45: 103939. [CrossRef]
Hsiao CH, Huang YL, Tse SL, et al. Automatic segment and quantify choroid layer in myopic eyes: deep learning based model. Semin Ophthalmol. 2022; 37: 611–618. [CrossRef] [PubMed]
Lim ZW, Li J, Wong D, et al. Comparison of manual and artificial intelligence-automated choroidal thickness segmentation of optical coherence tomography imaging in myopic adults. Eye Vis (Lond). 2024; 11(1): 21. [CrossRef] [PubMed]
Xu J, Xu L, Du KF, et al. Subfoveal choroidal thickness in diabetes and diabetic retinopathy. Ophthalmology. 2013; 120: 2023–2028. [CrossRef] [PubMed]
Zhang Q, Wang YX, Wei WB, Xu L, Jonas JB. Parapapillary Beta zone and gamma zone in a healthy population: the Beijing Eye Study 2011. Invest Ophthalmol Vis Sci. 2018; 59: 3320–3329. [CrossRef] [PubMed]
Shao L, Xu L, Chen CX, et al. Reproducibility of subfoveal choroidal thickness measurements with enhanced depth imaging by spectral-domain optical coherence tomography. Invest Ophthalmol Vis Sci. 2013; 54: 230–233. [CrossRef] [PubMed]
Chen J, Lu Y, Yu Q, et al. Transunet: Transformers make strong encoders for medical image segmentation. arXiv. Preprint posted online February 8, 2021, doi:2102.04306.
Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector. In: Computer Vision–ECCV2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 142016. Berlin: Springer International Publishing. (pp. 21–37).
Simonyan KJ . Very deep convolutional networks for large-scale image recognition. arXiv. Preprint posted online September 4, 2014, doi:1409.1556.
Kang SH, Hong SW, Im SK, Lee SH, Ahn MD. Effect of myopia on the thickness of the retinal nerve fiber layer measured by Cirrus HD optical coherence tomography. Invest Ophthalmol Vis Sci. 2010; 51: 4075–4083. [CrossRef] [PubMed]
Rao HL, Kumar AU, Babu JG, Kumar A, Senthil S, Garudadri CS. Predictors of normal optic nerve head, retinal nerve fiber layer, and macular parameters measured by spectral domain optical coherence tomography. Invest Ophthalmol Vis Sci. 2011; 52(2): 1103–1110. [CrossRef] [PubMed]
Ooto S, Hangai M, Yoshimura N. Effects of sex and age on the normal retinal and choroidal structures on optical coherence tomography. Curr Eye Res. 2015; 40: 213–225. [CrossRef] [PubMed]
Choma M, Sarunic M, Yang C, Izatt J. Sensitivity advantage of swept source and Fourier domain optical coherence tomography. Opt Express. 2003; 11: 2183–2189. [CrossRef] [PubMed]
Ostrin LA, Harb E, Nickla DL, et al. IMI-the dynamic choroid: new insights, challenges, and potential significance for human myopia. Invest Ophthalmol Vis Sci. 2023; 64(6): 4. [CrossRef] [PubMed]
Sadeghi E, Valsecchi N, Rahmanipour E, et al. Choroidal biomarkers in age-related macular degeneration. Surv Ophthalmol. 2025; 70: 167–183. [CrossRef] [PubMed]
Liu K, Li T, Zhong P, et al. Retinal and choroidal phenotypes across novel subtypes of type 2 diabetes mellitus. Am J Ophthalmol. 2025; 269: 205–215. [CrossRef] [PubMed]
Wang W, Liu S, Qiu Z, et al. Choroidal thickness in diabetes and diabetic retinopathy: a swept source OCT study. Invest Ophthalmol Vis Sci. 2020; 61(4): 29. [CrossRef] [PubMed]
Papathanasiou KA, Kazantzis D, Vrachatis DA, et al. Choroidal thickness in patients with systemic arterial hypertension: a systematic review and meta-analysis. Ther Adv Ophthalmol. 2022; 14: 25158414221132825. [PubMed]
Li JO, Liu H, Ting DSJ, et al. Digital technology, tele-medicine and artificial intelligence in ophthalmology: A global perspective. Prog Retin Eye Res. 2021; 82: 100900. [CrossRef] [PubMed]
Figure 1.
 
Automated choroidal segmentation. (A) OCT B-scan images that passed through the foveal center; (B) automated choroidal segmentation generated by TransUnet model.
Figure 1.
 
Automated choroidal segmentation. (A) OCT B-scan images that passed through the foveal center; (B) automated choroidal segmentation generated by TransUnet model.
Figure 2.
 
Scatter plot showing the correlation between the automated and manual subfoveal choroidal thickness measurements performed on optical coherence tomograms.
Figure 2.
 
Scatter plot showing the correlation between the automated and manual subfoveal choroidal thickness measurements performed on optical coherence tomograms.
Figure 3.
 
Bland–Altman plot demonstrating the agreement between automated and manual subfoveal choroidal thickness measurements.
Figure 3.
 
Bland–Altman plot demonstrating the agreement between automated and manual subfoveal choroidal thickness measurements.
Table 1.
 
Comparison of Baseline Characteristics Between Included and Excluded Participants, With Subgroup Analysis of Included Participants Stratified by Age and Axial Length
Table 1.
 
Comparison of Baseline Characteristics Between Included and Excluded Participants, With Subgroup Analysis of Included Participants Stratified by Age and Axial Length
Table 2.
 
Agreement and Correlation Between Manual and Automated Measurements of SFCT Stratified by Age, Axial Length, and SFCT
Table 2.
 
Agreement and Correlation Between Manual and Automated Measurements of SFCT Stratified by Age, Axial Length, and SFCT
×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×