Glaucoma is a major cause of irreversible blindness worldwide, resulting in the progressive degeneration of retinal ganglion cellss (RGCs), which makes early disease detection critical for effective management.Traditional methods for monitoring RGCs are labor-intensive and prone to errors. To address this, we propose Light-U-Net, a lightweight and scalable deep learning model designed to segment RGCs in retinal images. The model is trained and tested on a publicly available synthetic dataset as well as a self-generated real dataset. Additionally, we introduce a local maxima algorithm for counting RGCs based on generated annotations. A comparative analysis of various cell counting methods was performed, demonstrating that Light-U-Net combined with the watershed algorithm delivers superior segmentation and counting performance. These findings highlight the potential of Light-U-Net for automating RGC segmentation and counting, reducing errors, and improving efficiency compared to traditional approaches. Light-U-Net, in combination with watershed-based counting, provides an effective tool for glaucoma detection and monitoring, making it valuable for both disease progression tracking and treatment assessment. • Deep learning framework for automated retinal ganglion cells (RGC) counting in mouse retina frames. • Light-U-Net reduces computational cost while preserving accurate RGC detection. • Biologically informed preprocessing improves RGC centroid and region representation. • Evaluation on a self-generated RBPMS-labeled datasets across retinal regions and densities. • AI-based RGC counting enables reliable downstream retinal analysis in dense regions.
Gharaei et al. (Wed,) studied this question.