The thesis titled “Improving Performance of Convolutional Neural Networks” was presented by the master’s student Heba Kazem Dashr under the supervision of Professor Dr. Lamiaa Abdul Noor Mohammed.
In her study, the researcher applied a method to enhance recurrent neural networks in image-related tasks, specifically using pre-trained models NASNetMobile and Xeeption. The study aimed to address current limitations by providing a more efficient model for medical image applications. The proposed model was trained on various image windows to identify the best informative region. The proposed method was evaluated using three datasets: two for brain tumors in MRI scans, and a third for diabetic retinopathy. The proposed model achieved an accuracy of 99.83% on the first dataset and 99.57% on the second.

The achieved results demonstrated the superiority of the proposed model over existing ones in various technical metrics, indicating the success of this approach. The proposed method exhibited the ability to significantly enhance the performance and efficiency of CNN networks in medical image classification tasks, making it more effective for real-world applications. The thesis was accepted very good.