RetinaDetachNet: Automated Deep Learning Quantification of Photoreceptor Cell Death for Neuroprotection Studies in Experimental Retinal Detachment

Purpose: To develop and validate RetinaDetachNet, to our knowledge, the first validated deep learning pipeline for automated quantification of TUNEL-positive cells in experimental retinal detachment models. Methods: RetinaDetachNet combines a custom-trained U-Net for outer nuclear layer (ONL) segmentation with a hybrid approach for TUNEL-positive cell detection. StarDist provides initial nucleus segmentation; candidates are retained only if they satisfy area criteria and overlap sufficiently with Otsu-thresholded binary masks. Validation involved three independent datasets with temporal and institutional separation: primary (n = 50 images), historical (∼10 years prior; n = 50), and external (independent laboratory; n = 40). Agreement with manual counts (experienced and inexperienced observers) was assessed via Spearman correlation (5000 bootstrap iterations) and Bland-Altman analysis. Results: The U-Net achieved a Dice coefficient of 0.93 for ONL segmentation. RetinaDetachNet showed strong correlation with manual counting: Dataset 1, ρ = 0.87 (experienced) and ρ = 0.80 (inexperienced); Dataset 2, ρ = 0.98 (surpassing inter-observer ρ = 0.94); Dataset 3, ρ = 0.86 following calibration to local imaging parameters. The hybrid method outperformed StarDist-only (ρ = 0.70-0.84) and Otsu-only (ρ = 0.33-0.90) approaches across datasets. Bland-Altman analysis indicated minimal systematic bias. Conclusions: To our knowledge, RetinaDetachNet is the first validated deep learning pipeline for TUNEL-positive cell quantification in retinal detachment, delivering superior accuracy and reproducibility via its hybrid dual-validation architecture. Translational Relevance: This open-source tool, freely available on GitHub, enables standardized, observer-independent quantification of photoreceptor cell death in preclinical neuroprotection studies, reducing analysis time from hours to minutes and accelerating evaluation of candidate therapeutics.