كاربرد شبكه هاي عصبي مصنوعي و شبكه هاي عصبي كانولوشن در تشخيص تصاوير پزشكي: تمركز بر رتينوپاتي ديابتي

With the rapidly increasing global preva‎lence of diabetes mellitus, diabetic retinopathy (DR) has emerged as one of the leading causes of preventable blindness among working-age adults. Timely an‎d accurate severity grading of DR using retinal fundus images plays a critical role in determining appropriate treatment strategies an‎d preventing irreversible vision loss. However, conventional manual screening by ophthalmologists is time-consuming, costly, prone to inter-observer variability, an‎d is further exacerbated by the shortage of specialists, particularly in underserved regions. In recent years, deep learning systems—particularly convolutional neural networks (CNNs) an‎d transformer-based architectures—have shown considerable promise for the automated detection an‎d grading of diabetic retinopathy. Despite notable advances, several challenges persist, including the reliable detection of subtle lesions (such as microaneurysms, hemorrhages, an‎d hard exudates), severe class imbalance, visual similarity between adjacent severity stages, an‎d limited generalization across diverse datasets. These limitations hinder their widespread clinical adoption. This thesis presents an‎d eva‎luates a novel deep learning approach for the accurate grading of diabetic retinopathy. To this end, a simple three-layer convolutional neural network was systematically compared with six state-of-the-art pre-trained architectures: VGG16, ResNet50, ResNeXt50, ConvNeXt-Tiny, ViT-B/16, Swin-T, an‎d MNv4-Conv-L. The results of the comparative analysis indicate that the MobileNetV4-Conv-L model, achieving the highest accuracy rate (75%), demonstrated a significant superiority over the other models. This architecture achieved a performance improvement of approximately 20–23% compared with baseline an‎d older models, such as Simple-CNN (55% accuracy) an‎d VGG16 (52% accuracy). Furthermore, transformer-based architectures, such as ViT-B/16, with an accuracy of 59%, showed better performance than traditional convolutional neural networks. This quantitative improvement demonstrates that the application of modern pre-trained models in this field can provide acceptable performance while having the potential to become a valuable tool in the clinical diagnosis process. Such models can assist medical specialists in improving the accuracy of disease diagnosis.

محمدرضا احمدزاده , محمد داورپناه جزي

الهام محمودزاده , مينا اميري