Related Experiment Video
Updated: Dec 9, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.2K
Fully convolutional architecture vs sliding-window CNN for corneal endothelium cell segmentation
Juan P Vigueras-Guillén1,2, Busra Sari1, Stanley F Goes1
1Delft University of Technology, Dept. of Imaging Physics, Lorentzweg 1, Delft, 2628CJ The Netherlands.
BMC Biomedical Engineering
|September 9, 2020
Summary
This study introduces two convolutional neural networks (CNNs) for segmenting low-quality corneal endothelium (CE) images. The U-net model significantly improved accuracy and reduced errors in clinical parameter estimation compared to existing methods.
Area of Science:
- Ophthalmology
- Computer Vision
- Medical Imaging
Background:
- Corneal endothelium (CE) imaging is crucial for assessing corneal health.
- Automatic segmentation of CE images is challenging due to low image quality from in vivo techniques.
- Accurate segmentation is necessary for computing key clinical morphometric parameters.
Purpose of the Study:
- To develop and evaluate two novel convolutional neural networks (CNNs), U-net and a sliding-window network (SW-net), for segmenting corneal endothelium images.
- To assess the impact of probabilistic labels, contrast enhancement preprocessing, and Fourier analysis/watershed postprocessing on segmentation performance.
- To compare the performance of the proposed CNN methods against manual delineation and existing techniques.
Main Methods:
- Implementation of a global U-net based fully convolutional network and a local SW-net for CE image segmentation.
- Utilization of probabilistic labels instead of binary labels for improved segmentation accuracy.
- Application of a contrast enhancement preprocessing method and a postprocessing technique involving Fourier analysis and watershed transformation.
- Evaluation on 50 corneal endothelium images acquired using a Topcon SP-1P specular microscope.
Main Results:
- The U-net model achieved superior performance with a higher Area Under the Curve (AUC=0.9938) compared to SW-net (AUC=0.9921).
- Postprocessing resulted in a DICE score of 0.981 and a Modified Hausdorff Distance (MHD) of 0.22 for U-net, outperforming SW-net (DICE=0.978, MHD=0.30).
- U-net demonstrated statistically significant improvements in precision and accuracy for estimating cell density (ECD), polymegethism (CV), and pleomorphism (HEX), with minimal mean relative errors (0.4% for ECD, 2.8% for CV, 1.3% for HEX).
Conclusions:
- Both U-net and SW-net offer statistically significant advancements over current state-of-the-art methods for corneal endothelium segmentation.
- The U-net model achieved the lowest error rates and superior performance in clinical parameter estimation.
- Further improvements in segmentation performance can be achieved by refining the U-net approach based on prior work.
