Machine learning for segmenting cells in corneal endothelium images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Corneal endothelial cell analysis is crucial for assessing eye health.
• Current automated systems lack accuracy, while manual methods are time-consuming.
• Quantitative biomarkers like cell density and hexagonality are key indicators.

Purpose of the Study:

• To investigate and compare two deep learning methods, U-Net and SegNet, for automated cell segmentation in corneal endothelial images.
• To evaluate the performance of these deep learning models in accurately segmenting endothelial cells.

Main Methods:

• A dataset of 130 corneal endothelial images with expert-annotated cell borders was utilized.
• Two deep neural networks, U-Net and SegNet, were trained and tested for pixel-wise segmentation.
• Performance was evaluated using metrics such as Dice and Jaccard coefficients.

Main Results:

• The U-Net deep learning model demonstrated effective pixel-wise segmentation of corneal endothelial cells.
• Most segmentation errors produced by U-Net were visually insignificant.
• The U-Net approach shows potential for accurate cell segmentation and morphological analysis.

Conclusions:

• Deep learning, specifically the U-Net model, offers a viable solution for automated corneal endothelial cell segmentation.
• This automated analysis can lead to more efficient and accurate quantification of biomarkers for corneal health assessment.
• The U-Net method provides a promising tool for clinical applications in ophthalmology.