The scientific article focuses on the application of deep learning through simple U-Net, attention U-Net, residual U-Net, and residual attention U-Net models for diagnosing retinal diseases based on medical image analysis. The work includes a thorough analysis of each model's ability to detect retinal pathologies, taking into account their unique characteristics such as attention mechanisms and residual connections. The obtained experimental results confirm the high accuracy and reliability of the proposed models, emphasizing their potential as effective tools for automated diagnosis of retinal diseases based on medical images. This approach opens up new prospects for improving diagnostic procedures and increasing the efficiency of medical practice. The authors of the article propose an innovative method that can significantly facilitate the process of identifying retinal diseases, which is critical for early diagnosis and timely treatment. The results of the study support the prospect of using these models in clinical practice, highlighting their ability to accurately analyze medical images and improve the quality of eye health care.