Self-learning algorithm for denoising and advancing the integration of scRNA-seq datasets improves the identification of resilient and susceptible retinal ganglion cell types