Mobile-CellNet: Automatic Segmentation of Corneal Endothelium Using an Efficient Hybrid Deep Learning Model

Area of Science:

• Ophthalmology
• Biomedical Imaging
• Artificial Intelligence

Background:

• The corneal endothelium's hexagonal cell morphology and limited regeneration capacity make endothelial cell density (ECD) a key indicator of corneal health.
• Specular microscopy is crucial for imaging the corneal endothelium, but manual analysis for ECD measurement is labor-intensive and time-consuming.

Purpose of the Study:

• To develop and evaluate Mobile-CellNet, a novel, fully automated deep learning algorithm for efficient corneal endothelial cell segmentation and ECD estimation.
• To compare the performance of Mobile-CellNet against established U-Net and U-Net++ models in analyzing specular microscopy images.

Main Methods:

• Mobile-CellNet employs two parallel image segmentation models combined with classical image processing techniques for automated cell segmentation.
• The algorithm was benchmarked against U-Net and U-Net++ using a test dataset of specular corneal images.

Main Results:

• Mobile-CellNet achieved a mean absolute error of 4.06% for ECD estimation, comparable to U-Net's 3.80% error.
• The proposed algorithm demonstrated significantly higher efficiency, requiring 31 times fewer floating-point operations and 34 times fewer parameters than U-Net.

Conclusions:

• Mobile-CellNet provides accurate segmentation of corneal endothelial cells and efficient ECD and morphology analysis.
• This deep learning approach facilitates the development of tools for remote analysis of specular corneal endothelial images, improving accessibility and efficiency in clinical practice.