Comparative Evaluation of Deep Learning Architectures for Retinal Ganglion Cell Counting: FCRN-A, FCRN-A-v2, and U-Net

Deep Learning (DL) has revolutionized healthcare, particularly in disease prediction, medical imaging, and drug discovery. In ophthalmology, DL facilitates cell counting by detecting and quantifying retinal ganglion cells (RGCs) from microscopy images. Manual counting is labor-intensive and errorprone, leading to the development of automated DL systems. This paper compares three architectures, namely, Fully Convolutional Regression Network FCRN-A, FCRN-A-v2, and U-Net, using a synthetic dataset and a custom real dataset. The networks, trained through supervised learning, produce density maps, with model performance evaluated via 5-fold cross-validation based on mean absolute error (MAE) and standard deviation (STD). Results show U-Net outperforms the others, achieving the lowest MAE and STD in cell counting.