Brain Tumor Image Classification using Fine-tuned SqueezeNet Model

Abstract

Brain tumor detection and classification using deep learning techniques have shown great potential in assisting with early diagnosis, providing clinical guidance, and improving treatment effectiveness in neurology. However, existing studies have overlooked important factors and suffered from increased computation time. In this research, we address these limitations by leveraging a broader range of brain Medical Record images (MRI) with varying weights and contrast enhancement methods. Our chosen model, SqueezeNet, tackles these challenges effectively. Using a dataset comprising various modalities of MR images obtained from the University of Gondar Specialized Hospital, including both normal brain scans and scans depicting different tumor conditions, we employ the SqueezeNet model for accurate classification. Preprocessing techniques, such as converting DICOM to JPG and applying filtering and segmentation methods, enhance performance. Our proposed model achieves a SoftMax classification accuracy of 97.98% while maintaining computational efficiency with a computation time of 13,868.13 seconds. Compared to other evaluated models, SqueezeNet exhibits the shortest computation time, surpassing the accuracy of models like VGGNet19 and AlexNet. Furthermore, we explore concatenation models, and the combination of SqueezeNet with VGG19 demonstrates the highest accuracy of 97.36%. In summary, our research contributes to the advancement of brain tumor detection and classification by addressing previous limitations and achieving high accuracy and computational efficiency. Future studies should focus on integrating tumor grading, exploring a broad