Abstarct: Abstract Glaucoma is one of the leading causes of irreversible vision loss, associated with damage to the optic nerve. Early detection of this disease is crucial to prevent its progression and preserve vision. However, due to the asymptomatic nature of glaucoma in its early stages and the variability of related medical images, accurate and timely diagnosis remains a significant challenge. Additionally, issues such as data imbalance and low-quality medical images in real-world environments hinder the effectiveness of deep learning models in this domain. This study explores and develops novel deep learning approaches for glaucoma diagnosis through medical image analysis. In the first part, data from Optical Coherence Tomography (OCT) in the Shahroud Eye Cohort dataset with imbalanced data were utilized. To address the challenge of data imbalance, a hybrid method baxsed on weighted ensemble learning and data augmentation techniques was introduced. This method, by augmenting glaucoma-related data sixfold and dividing normal data into five different groups, helped balance the data and improve model performance. Furthermore, the use of advanced architectures such as ResNet, multi-scale feature extraction, cross-attention mechanisms, and channel and spatial attention modules led to a significant improvement in model accuracy. As a result, this method achieved an accuracy of 98.90% in glaucoma detection, showing a 2.2% improvement compared to the previous model (ConvNeXtLarge). In the second part, a novel approach for processing retinal images using deep neural networks and a multi-scale attention block architecture was introduced. This architecture can extract important image features at different scales and prevent overfitting, leading to increased diagnostic accuracy. Experiments on the ACRIMA dataset showed that this method achieved an accuracy of 97.18%, surpassing previous results. This study, by presenting innovative methods in processing and analyzing medical images, offers hope for improving glaucoma diagnosis and addressing the challenges related to data imbalance and image quality in real-world settings.