زهرا داورزني

Eye diseases are among the leading causes of visual impairment an‎d blindness worldwide, an‎d early diagnosis plays a crucial role in preventing their progression. With advancements in image processing an‎d deep learning technologies, it has become possible to automatically analyze medical images an‎d detect diseases with high accuracy. Convolutional Neural Networks (CNNs), known for their ability to extract local features, have shown strong performance in classifying eye diseases. However, they face limitations in capturing long-range dependencies across different regions of the retina. Transformers, originally developed for natural language processing, have emerged as a powerful architecture for image classification. By dividing an image into small patches an‎d applying self-attention mechanisms, Vision Transformers (ViTs) are capable of modeling long-distance relationships within the image. Despite their strengths, ViTs are highly dependent on the quality an‎d size of the training datasets. To overcome the individual limitations of CNNs an‎d ViTs, this study proposes a hybrid model that integrates both architectures. By leveraging the complementary strengths of CNNs (for local feature extraction) an‎d ViTs (for long-range dependency modeling), the proposed model achieves improved accuracy an‎d efficiency in retinal image analysis, particularly in cases with subtle an‎d scattered structural changes. The model reached an accuracy of 0.9969, outperforming several state-of-the-art deep learning models.

هوش مصنوعي , پردازش تصوير , يادگيري عميق , پيش¬پردازش , استخراج ويژگي , شبكه‌هاي عصبي كانولوشني , ترنسفورمر بينايي , مكانيزم خودتوجهي

Artificial Intelligence , Image Processing , Deep Learning , Preprocessing , Feature Extraction , Convolutional Neural Networks (CNNs) , Vision Transformer (ViT) , Self-Attention Mechanisms

Zahra Davarzani

Dr. Seyed Mohammadreza Mousavi mirkolaei