Automatic segmentation of calibration device boundaries on ophthalmic optical biometry instruments in optical coherence tomography images

Introduction

In the current landscape of ophthalmology, advancements in optical biometry persistently redefine our understanding of ocular structure and functionality. More specifically, in the accurate assessment of axial length (AL) (1-4), this technology plays a crucial role in refractive surgery planning, intraocular pressure monitoring, and the early detection of diverse ocular pathologies like cataracts (5-7). Minute fluctuations in AL are critical in minimizing unfavorable visual consequences post-refractive surgery. However, due to the proliferation and escalating complexity of optical biometry devices in the market (8-10), conventional measurement methodologies are facing substantial technical impediments in attaining high-precision calibration (11-14).

Optical coherence tomography (OCT) is primarily classified as time-domain OCT and Fourier-domain optical coherence tomography (FD-OCT) (15-17), the latter of which is further divided into spectral domain optical coherence tomography (SD-OCT) and swept-source OCT (18,19). SD-OCT is a non-invasive imaging technique that is widely applied in medical and biomedical domains, particularly in ophthalmology (20-23). This technology generates high-resolution cross-sectional tissue images by assessing light interference. The SD-OCT system uses a laser light source, splitting the light beam into reference and sample paths. This interference phenomenon supplies insights into light path disparities. The system employs a spectrometer to analyze the interference signal’s spectral components, consequently gauging the disparities in light paths. Ultimately, the SD-OCT system translates these data into high-resolution tissue images, unveiling internal sample structures like the eye’s retina. Additionally, SD-OCT can generate three-dimensional images showcasing the sample’s volumetric structure.

In the calibration of ophthalmic optical biometers, the intrinsic performance of the system is of paramount importance, as the stability of the system underpins the subsequent clinical measurements of various parameters such as AL and corneal curvature. However, due to inherent variations among different machine systems, it is essential to calibrate the biometer to ensure precise clinical measurements. Typically, the calibration of the biometer involves the use of glass rods with fixed group refractive indices to simulate the model from the anterior to the posterior segment of the human eye. Additionally, during the calibration process, it is crucial to obtain the pixel heights (PHs) of the glass rod boundaries in OCT images, as the accuracy of these PHs directly affects the final calibration outcome. Numerous segmentation algorithms have been applied to the segmentation of foreground targets in OCT images; however, currently, there is no method specifically tailored for the boundary localization of calibration devices in the context of biometer calibration. While deep-learning and other advanced methods are currently popular for segmentation, training a model specifically for the calibration device in this scenario would be a tedious process and has not been deemed necessary. Therefore, the integration of traditional image processing algorithms with mathematical statistical methods presents a significant advantage in this context.

Thus, this study sought to establish a pioneering algorithm targeting the enhancement of boundary detection precision in SD-OCT images to calibrate ophthalmic optical biometers (i.e., the K9 standard, the convex surface, and an AL ranging from 9.987 to 34.728 mm) by employing advanced image processing methodologies. The front surface of the calibration device is akin to the anterior segment of the human eye, and its rear surface resembles the posterior segment of the eye. The methodology employs sophisticated image morphological techniques to extract and enhance the calibration device’s edge information from OCT images. Additionally, a clustering analysis-driven strategy is used to meticulously refine and optimize the extracted boundary points, thus achieving highly precise localization of boundary vertex PHs. Through rigorous assessment of the algorithm’s repeatability and recognition accuracy, we validated its effectiveness and reliability in clinical use. This pioneering study not only addresses an issue in current technology but also provides more accurate data support for ophthalmic diagnosis and treatment, which could elevate the caliber and efficacy of ophthalmic healthcare services.

Methods

In this study, the algorithm was formulated in the Visual Studio 2022 integrated development environment (released in 2021 by Microsoft, USA), installed on a Windows 10-operated personal computer. The axial imaging data were captured using the BVB2000 (Big Vision, Suzhou, China) system. The BVB2000 acquires OCT images that include both the anterior and posterior segments of the eye. In these images, the left side corresponds to the anterior segment, while the right side corresponds to the posterior segment. The grayscale image obtained boasts a resolution of 1,180 by 1,024 pixels. The obtained images were exported and stored in bitmap format for subsequent analysis employing our algorithm. The formulated algorithm employs the following two-step process: (I) the detection of boundary vertices for the anterior and posterior sections of the calibration device; and (II) the determination of the PHs for the detected boundary vertices.

Figure 1 (original image) provides a schematic flowchart of the procedures involved in the calibration device anterior and posterior segment extraction algorithm, featuring an exemplar axial imaging depiction. To automatically discern the locations of the anterior and posterior segments, the original image is bifurcated into two parts. This division is accomplished through Gaussian fitting applied to the mean intensity along the x-axis of the original image. The index of the second peak serves as the segmentation threshold (Figure 1, I_A and I_P).

Figure 1 The algorithm for positioning the anterior segment and posterior segment of the calibration device. I_A, anterior OCT image of the calibration device; I_FM, fitting diagram of upper and lower boundary points in anterior of the calibration device; I_KO, anterior information extraction diagram of calibration device; I_KF, posterior information extraction diagram of calibration device; I_LAB, the final fitting result graph of this algorithm; I_P, posterior OCT image of the calibration device; I_PF, fitting diagram of upper and lower boundary points in posterior of the calibration device; OCT, optical coherence tomography.

To ascertain the y-directional position of the anterior segment vertex in I_A for the calibration device, the initial step involves binarizing the image to produce a binary mask representation (24). Given the substantial noise at the image’s boundary, a square structural element sized at 1,180 by 20 pixels is used to exclude the image border. Additionally, pixel blocks smaller than 5 pixels are eliminated. Simultaneously, to emphasize object edges, the resultant image undergoes a logical AND operation with I_A (Figure 1, I_A) and is subsequently processed using a Sobel operator (Figure 1, I_KO). Additionally, small voids are present within the connected components of the image, where all holes with an area smaller than 1,000 pixels within these regions are subsequently filled (25). Following image acquisition, a top-down scanning technique is used to extract the initial and terminal pixel points with a value of 255 in every column. Subsequently, the remaining pixel points are removed, and only the pertinent pixel points are retained. Due to the discrete nature of the extracted pixel points, the following multistep procedure is employed: (I) the points undergo clustering via mean shift with limited iterations (26), resulting in n point clusters; (II) the minimum bounding rectangle area for each cluster is computed, discarding clusters larger than 350 or smaller than 80 in the area; and (III) a quadratic polynomial regression model is employed to perform curve fitting for all points within the processed clusters. The resultant fitted curve, depicted as the blue and green lines in Figure 1, I_FM, delineates the correlation between the extracted pixel points and their respective positions. Following this, the coordinates of the fitted curve’s vertex are recorded, followed by meticulous mapping to correlate fitting line with the corresponding position in the original image (Figure 1, I_LAB).

To extract the y-directional position of the posterior segment in I_P, mirroring the methodology used for the anterior segment, the image undergoes a sequential process involving binarization, boundary elimination, preservation of the most extensive connected component, and hole filling (Figure 1, I_KF). Upon image acquisition, the upper and lower boundaries of the posterior segment are extracted using a method identical to that used for the anterior segment’s boundary point extraction. This is followed by distinct quadratic polynomial fittings, depicted as the red and yellow lines in Figure 1, I_PF, elucidating the connection between the extracted pixel blocks and their respective positions. Afterwards, the fitted curve’s vertex coordinates are recorded, followed by meticulous mapping to correlate it with the corresponding position in the original image (Figure 1, I_LAB).

After acquiring the boundary vertices for both the anterior and posterior segments of the calibration device, the PH is determined by calculating the vertical distance from the vertex coordinate to the lower boundary of the image. After the computation of the PHs, PHs undergo preparatory steps for subsequent analysis and processing.

Results

To assess the efficacy of this automated algorithm, a set of 15 axial imaging instances capturing the ophthalmic optical biometer calibration device with varying ALs was captured using the BVB2000 (Big Vision) system. The calibration of all the ophthalmic optical biometer devices was performed at the China Institute of Metrology. Devices with axial measurements of 9.987, 15.042, 20.011, and 30.024 mm had an uncertainty measurement of 0.010 mm, documented under report number YXgg2023-00725. Similarly, devices measuring 10.025, 13.370, 17.823, 26.694, and 34.728 mm had an uncertainty measurement of 0.005 mm, detailed under report number YXgg2023-02243. The algorithm autonomously processed the axial imaging data and conducted the PH calculations, processing 15 images in approximately 0.5 seconds.

Performance validation for identification repeatability

Figure 2 displays individual representations of ophthalmic optical biometer calibration devices distinguished by varying ALs, identified by an automated algorithm. Figure 2A-2I show boundary identification depictions corresponding to calibration devices measuring 9.987, 10.025, 13.370, 15.042, 17.823, 20.011, 26.694, 30.024, and 34.728 mm, respectively. In these images, the green curve symbolizes the fitted lower boundary of the anterior segment, while the red curve signifies the fitted upper boundary of the posterior segment in the calibration device. The documented coordinates of the fitted curve vertices facilitate subsequent computational analyses. To ascertain the automated algorithm’s consistency in delineating the anterior and posterior segment boundaries of the calibration devices in the images, 15 real-time captures were separately conducted on nine devices with varied ALs. The repeatability of the anterior and posterior segment PHs, denoted as Spixel, was determined using the following formula:

Spixel=∑i=115(Dpixel¯−Dpixel-i)214[1]

Figure 2 The identification results of the anterior and posterior segments of nine representative ophthalmic optical biometer calibration devices with varying sizes. The display sequence from (A-I) is based on their sizes, which range from 9.987 to 34.728 mm.

where Dpixel¯ is the mean PH derived from the results of the 15 identifications, and Dpixel-i is the PH obtained from the i recognition. The repeatability of both the anterior and posterior segment PHs, derived from 15 real-time identifications using an automated algorithm across nine distinct AL calibration devices, is summarized in Table 1. Notably, the device with an AL of 9.987 mm had a repeatability of 0.238 for anterior segment PH identification and 0.171 for posterior segment PH identification. Similarly, the device measuring 10.025 mm had a repeatability of 0.124 for both the anterior and posterior segment PHs. This observed pattern persisted across other devices: 13.370 mm (anterior: 0, posterior: 0), 15.042 mm (anterior: 0, posterior: 0), 17.823 mm (anterior: 0.067, posterior: 0.124), 20.011 mm (anterior: 0.210, posterior: 0.067), 26.694 mm (anterior: 0.238, posterior: 0), 30.024 mm (anterior: 0.067, posterior: 0), and 34.728 mm (anterior: 0.352, posterior: 0).

Performance validation of recognition accuracy

While variations in the sample arm optical path of the OCT imaging system can influence the image heights of the calibration device, different optical path lengths (OPLs) are anticipated to demonstrate a linear trend in variation with the boundary PHs at their respective positions (16). To further corroborate the algorithm's precision in identifying the calibration device's boundary PHs, an instance with an AL of 15.042 mm is specifically examined. Initially, the algorithm undergoes real-time identification of both the anterior and posterior segments of the calibration device across various optical path positions, repeated 15 times each. Subsequently, the averaged PHs for both segments across the 15 identifications at each optical path position are computed. Table 2 comprehensively delineates the averaged PHs of the lower boundary vertex in the anterior section and the upper boundary vertex in the posterior section across 15 real-time captures of the 15.042 mm axial length calibration device, spanning various OPLs in the sample arm.

Finally, a linear regression analysis was conducted between the averaged anterior and posterior segment PHs and the OPL where the calibration device was positioned. Regarding the relationship between the PH of the lower boundary vertex of the anterior section and the sample arm OPL, the resulting one-variable linear fitting formula was expressed as:

0.0066xAP+yAO−45.6755=0[2]

where y_AO represents the sample arm OPL, and x_AP represents the average PH of the lower boundary vertex at different OPLs. Similarly, regarding the relationship between the PH of the upper boundary vertex of the posterior section and the sample arm OPL, the obtained one-variable linear fitting formula was expressed as:

0.0034xPP+yPO−20.9615=0[3]

where y_PO represents the sample arm OPL, and x_PP represents the average PH of the upper boundary vertex at various OPLs. Figure 3 shows the linear regression performed with the anterior section’s lower boundary vertex PH as the independent variable and the OPL in the sample arm as the dependent variable. Meanwhile, Figure 4 shows the linear regression conducted using the posterior section’s upper boundary vertex PH as the independent variable and the OPL in the sample arm as the dependent variable. After obtaining the fitting results, the precision of the anterior and posterior segment PHs of the recognized calibration device through this algorithm was evaluated by computing its coefficient of determination (R²) (27). The coefficient of determination was calculated using the following formula:

R2=1−∑i=1n(Xi−Yi)2∑i=1n(1n∑i=1nYi−Y¯)2[4]

Figure 3 Linear regression plot showing the relationship between the pixel height of the anterior boundary vertex of the calibration device and the corresponding sample arm optical path length.

Figure 4 Linear regression plot showing the relationship between the pixel height of the posterior boundary vertex of the calibration device and the corresponding sample arm optical path length.

where X_i is the ith predicted value, and Y_i is the ith actual value. The R² ranges between (−∞, 1) where a value of 1 indicates the best fit and −∞ indicates the poorest fit. R² represents the proportion of variance in the dependent variable that is predictable from the independent variable. The R-squared values for the actual and fitted values of the anterior and posterior segments, calculated using the above formulas, were 0.9998330 and 0.9999863, respectively. The obtained two determination coefficients were very close to the optimal value of 1, which indicates a fully linear relationship between the PHs of the boundary vertices of the anterior and posterior segments of the calibrated device, as located by this algorithm, and the OPL in the sample arm.

Discussion

This study successfully developed an automated algorithm specifically designed to accurately identify the PH of calibration device boundary vertices in ophthalmic optical biometry instruments. By automatically segmenting the calibration device boundaries in OCT images, this algorithm not only improves measurement accuracy but also provides more accurate data support for ophthalmic diagnosis and treatment. In addition, the developed algorithm overcomes the technical barriers encountered by traditional measurement methods in high-precision calibration, demonstrating the potential application of advanced image processing technology in the field of ophthalmic optical measurement.

The innovation of the algorithm lies in its comprehensive application of image morphology processing technology and clustering analysis to accurately extract and optimize the boundary points of the calibration device in OCT images (26). The successful implementation of this method not only shows the application value of advanced image processing technology in ophthalmic optical measurement, but also reveals the key technical problems that need to be solved in practical applications. Specifically, the challenges faced by algorithms include how to effectively handle noise in images, and how to maintain high repeatability and recognition accuracy on different models and brands of optical biometric instruments.

The potential of algorithms to improve the calibration accuracy of ophthalmic optical biometric instruments is of great significance in enhancing the quality of ophthalmic medical services. Accurate calibration is the foundation for effective ophthalmic diagnosis and treatment, especially in refractive surgery planning, intraocular pressure monitoring, and the early detection of eye diseases such as cataracts (1,3,5). Therefore, the algorithm proposed in this study can provide ophthalmologists with more reliable data, helping them make more accurate diagnoses and treatment decisions.

This study has achieved positive results; however, future research will need to address various issues. First, the universality and adaptability of the algorithm need to be validated in a wider range of application scenarios, including optical biological measurement instruments of different types and brands. Second, further research should seek to optimize the processing speed and user interface of algorithms to improve their usability and acceptance in clinical environments. In addition, considering the rapid development of artificial intelligence and machine-learning technologies in the field of medical image analysis, future work should explore the combination of these technologies with current algorithms to further improve their accuracy, efficiency, and automation level (16,23).

Therefore, the automation algorithm developed in this study represents a technological advancement in the field of ophthalmic optical biometry, and its potential to improve measurement accuracy and support ophthalmic diagnosis and treatment deserves further exploration and development. Future research will be crucial, not only in optimizing and validating existing algorithms, but also in exploring new technological integration paths to continuously promote the development and innovation of ophthalmic medical technology.

Conclusions

This study successfully developed an automation algorithm that employs an innovative method to precisely identify calibration device boundary vertices in ophthalmic optical biometry instruments. By applying advanced image processing techniques, this algorithm significantly improves the accuracy and reliability of measurements, providing strong data support for ophthalmic diagnoses and treatment decisions. Our experimental results showed that the algorithm has high repeatability and recognition accuracy, demonstrating its enormous potential in clinical applications. Future research will explore the widespread application of algorithms, aiming to continuously improve the quality and efficiency of ophthalmic medical services through technological innovation. This algorithm may provide a novel, rapid, and straightforward approach for segmenting the boundaries of calibration devices in the context of biometer AL calibration. Our findings provide valuable insights into the segmentation of calibration device images acquired by other OCT systems. Given that most biometer AL calibration devices exhibit nearly identical representations in OCT images, readers could achieve satisfactory results by making minor adjustments to the workflow of the proposed method when replicating this algorithm.

Acknowledgments

None.

Footnote

Funding: This study was supported in part by the Natural Science Foundation of the Jiangsu Higher Education Institutions of China (No. 24KJB510042), the National Nature Science Foundation of China (Nos. 62271337, 62371328, and 62371326), and the Natural Science Foundation of Jiangsu Province of China (Nos. BK20211308 and BK20231310).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-2051/coif). F.C. is an employee of Suzhou Big Vision Medical Imaging Technology Co., Ltd.; however, no conflicts of interest arose in relation to the content of this article and his affiliation. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Khorrami-Nejad M, Khodair AM, Khodaparast M, Babapour Mofrad F, Dehghanian Nasrabadi F. Comparison of the ocular ultrasonic and optical biometry devices in the different quality measurements. J Optom 2023;16:284-95. [Crossref] [PubMed]
2. Nagayama M, Kimura S, Hosokawa MM, Shiode Y, Matoba R, Morita T, Kanenaga K, Morizane Y. Comparative analysis of axial length measurement method for eyes with submacular hemorrhage. Jpn J Ophthalmol 2025;69:196-202. [Crossref] [PubMed]
3. Teshigawara T, Meguro A, Mizuki N. Relationship Between Postoperative Intraocular Lens Shift and Postoperative Refraction Change in Cataract Surgery Using Three Different Types of Intraocular Lenses. Ophthalmol Ther 2021;10:989-1002. [Crossref] [PubMed]
4. Lee CY, Jeng YT, Yang SF, Huang CT, Chao CC, Lian IB, Huang JY, Chang CK. Topographic and Surgical Risk Factors for Early Myopic Regression between Small Incision Lenticule Extraction and Laser In Situ Keratomileusis. Diagnostics (Basel) 2024;14:1275. [Crossref] [PubMed]
5. Lwowski C, Kaiser KP, Bucur J, Schicho P, Kohnen T. Accuracy of using the axial length of the fellow eye for IOL calculation in retinal detachment eyes undergoing silicone oil removal. Br J Ophthalmol 2024;108:921-6. [Crossref] [PubMed]
6. Chonpimai P, Chirapapaisan C, Srivannaboon S, Loket S, Nujoi W, Dongngam S. Double peak axial length measurement signal in cataract patients with epiretinal membrane. Int Ophthalmol 2023;43:1337-43. [Crossref] [PubMed]
7. Oh J, Kim YH, Yun C, Ahn SM. Comparison of axial length measurement in eyes with chorioretinal diseases using spectral-domain optical coherence tomography-based device and partial coherence interferometry. Investigative Ophthalmology & Visual Science 2022;63:3077-F0549.
8. Ominde BS, Abadom GE, Ikubor JE, Achapu LC, Igbigbi PS. Normal Diameters of Extraocular Muscles: A Nigerian Retrospective Study. Niger Postgrad Med J 2024;31:147-55. [Crossref] [PubMed]
9. Gohari M, Babaei E, Attar A, Shiravani M, Salehi L. Change in Biometric Parameters after Deep Vitrectomy with Silicon Oil Tamponade. Pakistan Journal of Ophthalmology 2024;40: [Crossref]
10. Boily L, Michaud L, Garon ML, Marcotte R. Effects of Optical Zone Variation of High-Addition Multifocal Contact Lenses on the Global Flash Multifocal Electroretinography. Eye Contact Lens 2024;50:315-20. [Crossref] [PubMed]
11. Kolosky T, Das U, Panchal B, Byun S, Dolgetta A, Levin MR, Alexander JL. Anterior Chamber Depth and Lens Thickness Measurements in Pediatric Eyes: Ultrasound Biomicroscopy Versus Immersion A-Scan Ultrasonography. Ultrasound Med Biol 2024;50:1346-51. [Crossref] [PubMed]
12. Lau J, Koh WL, Ng JS, Lee D, Peh CH, Lam J, Tan KK, Koh V. How can we better evaluate paediatric progression of myopia and associated risk factors? Lessons from the COVID-19 pandemic: A systematic review. Acta Ophthalmol 2024;102:e257-71. [Crossref] [PubMed]
13. Mahmoud A, Pomar L, Lambert V, Picone O, Hcini N. Prenatal and Postnatal Ocular Abnormalities Following Congenital Zika Virus Infections: A Systematic Review. Ocul Immunol Inflamm 2024;32:2217-27. [Crossref] [PubMed]
14. Shao Y, Ma JM, Huang XM. Guidelines for standard operation of imaging modalities in orbital diseases (2024). Int J Ophthalmol 2025;18:51-66. [Crossref] [PubMed]
15. Kim H, Yun H, Jeong S, Lee S, Cho E, Rho J. Optical Metasurfaces for Biomedical Imaging and Sensing. ACS Nano 2025;19:3085-114. [Crossref] [PubMed]
16. Mokhtari A, Maris BM, Fiorini P. A Survey on Optical Coherence Tomography—Technology and Application. Bioengineering 2025;12:65. [Crossref] [PubMed]
17. Bikbov MM, Kazakbaeva GM, Panda-Jonas S, Mustafina GR, Jonas JB. Choroidal thickness under pilocarpine versus cyclopentolate. Sci Rep 2025;15:2221. [Crossref] [PubMed]
18. Yang S, Xin Z, Cheng W, Zhong P, Liu R, Zhu Z, Zhu LZ, Shang X, Chen S, Huang W, Zhang L, Wang W. Photoreceptor metabolic window unveils eye-body interactions. Nat Commun 2025;16:697. [Crossref] [PubMed]
19. Naveen Kumar P, Koilpillai RD, Bhattacharya S. Enhanced A-scan spatial resolution in spectral domain OCT exploiting the Wigner-Ville technique. Optics and Lasers in Engineering 2025;186:108736.
20. Fujii R, Matsushita M, Itani Y, Hama A, Natsume T, Takamatsu H. Intravitreal Administration of Avacincaptad Pegol in a Nonhuman Primate Model of Dry Age-Related Macular Degeneration. Pharmacol Res Perspect 2025;13:e70052. [Crossref] [PubMed]
21. Truong GB, Tran TT, Than NL, Nguyen VQ, Nguyen TH, Pham VT. SC-MambaFew: Few-shot learning based on Mamba and selective spatial-channel attention for bearing fault diagnosis. Computers and Electrical Engineering 2025;123:110004.
22. Wang Y, Bai L. Accurate Monte Carlo simulation of frequency-domain optical coherence tomography. Int J Numer Method Biomed Eng 2019;35:e3177. [Crossref] [PubMed]
23. Abira Bright B, Lakshmi Parvathi M, Damodaran V. Simulated retinal layers to study optical coherence tomography imaging. Women in Optics and Photonics in India 2022. SPIE; 2023.
24. Parlak İE, Emel E. Deep learning-based detection of internal defect types and their grades in high-pressure aluminum castings. Measurement 2025;242:116119.
25. Jähne B. Digital image processing. Springer Berlin, Heidelberg; 2005.
26. Lu C, Xia S, Shao M, Fu Y. Arc-support Line Segments Revisited: An Efficient High-quality Ellipse Detection. IEEE Trans Image Process 2019; Epub ahead of print. [Crossref]
27. Li A, Chu J, Huang S, Liu Y, He M, Liu X. Machine learning-assisted development of gas separation membranes: A review. Carbon Capture Science & Technology 2025;14:100374.