EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

RefraSegNet: OCT Image Segmentation Framework for Refraction Correction and Accurate Biometric Measurements of the Eye

Muhammad Usman, Keerthana Soman, Majad Mansoor, Ireneusz Grulkowski, Jacek Ruminski

IEEE Access · 2026 · 14 · 128202-128221

DOI: 10.1109/access.2026.3717540 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Muhammad Usman und 4 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 60 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: European Union’s Horizon Europe research and innovation program, Marie Skłodowska-Curie Actions (MSCA) Doctoral Network (DN), Horizon Europe (HORIZON)-Marie Skłodowska-Curie Actions (MSCA)-2022-DN, “Improving BiomEdical diagnosis through LIGHT-based technologies and machine learning (BE-LIGHT)”.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate quantitative analysis of anterior segment optical coherence tomography (AS-OCT) images requires both pixel-level boundary delineation and correction of refractive distortions inherent to optical path length measurements. This study presents RefraSegNet, a refractive-aware deep learning framework that integrates multi-class segmentation with physics-based ray tracing for geometrically consistent biometric assessment of the cornea and crystalline lens, including the lens nucleus. The proposed architecture combines a pre-trained ResNet encoder, a multi-scale feature refinement module, and an attention-based integration mechanism to enhance boundary continuity and robustness under speckle noise, low contrast, and shadow artifacts. To mitigate annotation bias and improve generalization, a fuzzy-label training strategy is introduced. The model was trained and validated on long-range OCT scans from 60 subjects ( $26 \pm 7$ years) across different accommodation demands. RefraSegNet achieved a mean intersection-over-union (mIoU) of 0.936 for multi-structure segmentation with a recall of 0.097 and a precision of 0.995 for ocular structures’ segmentation. After refraction correction, quantitative geometric analysis yielded an overall mean absolute error of 0.031 mm and mean percentage error of 2.5% in thickness measurements. Additionally, the measurements of radii of curvature of anterior segment surfaces were characterized by an overall mean absolute error of 0.444 mm, and mean percentage error of 5.6%. The proposed approach advances AS-OCT toward robust, geometry-aware, and clinically relevant quantitative assessment. The proposed framework provides a fully automated, refraction-aware pipeline that translates raw OCT images into quantitative biometric parameters.

Förderung

  • European Union’s Horizon Europe research and innovation program
  • Marie Skłodowska-Curie Actions (MSCA) Doctoral Network (DN), Horizon Europe (HORIZON)-Marie Skłodowska-Curie Actions (MSCA)-2022-DN
  • “Improving BiomEdical diagnosis through LIGHT-based technologies and machine learning (BE-LIGHT)”

Zitieren

APA

Usman, M., Soman, K., Mansoor, M., Grulkowski, I., & Ruminski, J. (2026). RefraSegNet: OCT Image Segmentation Framework for Refraction Correction and Accurate Biometric Measurements of the Eye. IEEE Access, 14, 128202-128221. https://doi.org/10.1109/access.2026.3717540

Vancouver

Usman M, Soman K, Mansoor M, Grulkowski I, Ruminski J. RefraSegNet: OCT Image Segmentation Framework for Refraction Correction and Accurate Biometric Measurements of the Eye. IEEE Access. 2026;14:128202-128221. doi:10.1109/access.2026.3717540

BibTeX
@article{Usman2026,
  author = {Usman, Muhammad and Soman, Keerthana and Mansoor, Majad and Grulkowski, Ireneusz and Ruminski, Jacek},
  title = {RefraSegNet: OCT Image Segmentation Framework for Refraction Correction and Accurate Biometric Measurements of the Eye},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {128202--128221},
  doi = {10.1109/access.2026.3717540}
}
Kostenlos starten →