Related Experiment Video
Updated: Mar 14, 2026
Novel Photoacoustic Microscopy and Optical Coherence Tomography Dual-modality Chorioretinal Imaging in Living Rabbit Eyes
Published on: February 8, 2018
Automatic deep learning-based segmentation of cornea and lens in 2D OCT images of rabbit eyes
Wannes De Martelaere1, Haijun Lv2, Hanrui Li2
1Britton Chance Center for Biomedical Photonics, Wuhan National Laboratory for Optoelectronics, Huazhong University of Science and Technology, Wuhan, 430074, Hubei, China; MOE Key Laboratory for Biomedical Photonics, Department of Biomedical Engineering, College of Life Science and Technology, Advanced Biomedical Imaging Facility, Huazhong University of Science and Technology, Wuhan, 430074, Hubei, China; Vrije Universiteit Brussel, Department of Photonics Engineering, Pleinlaan 2, 1050, Brussels, Brussels-Capital Region, Belgium; Korea University, Department of Nonlinear Dynamics and Biophysics, 145 Anam-ro, Seongbuk-gu, Seoul, 02841, Seoul, South Korea.
Abstract:
Anterior-segment optical coherence tomography (AS-OCT) supports cataract surgery planning by revealing corneal and crystalline lens geometry, but clinical workflows require fast, user-independent delineation of key interfaces in noisy scans. We address this need with an end-to-end pipeline tailored to low signal to noise ratio (SNR), preclinical AS-OCT: a systematic architecture search selects a U-Net with an ImageNet-pretrained EfficientNet-B2 encoder; a residual-driven, mask-guided non-local means stage suppresses background clutter while preserving boundaries; and a mask-to-surface conversion fits robust third-order polynomials to produce biometry-ready corneal and lenticular interfaces. Trained on 440 B-scans of rabbit eyes from a custom system, the model achieves mean Intersection-over-Union (IoU) of 0.957 ± 0.007 (cornea) and 0.978 ± 0.007 (lens) across five validation splits, with boundary root-mean-squared-error (RMSE) of 4.03-6.91 μ m between the predicted and manual reference surfaces. On an operator/time/subject-independent test set (n = 80), performance remains high-IoU 0.939 ± 0.008 (cornea) and 0.961 ± 0.009 (lens)-and the method outperforms a classical graph-search baseline. Despite limited training data and challenging image quality, the approach delivers accurate, robust surfaces suitable for downstream biometry.
