Efficient Feature Learning Based Xception CNN Model Optimization for MRI Brain Tumor Image Classification

• Published in: Traitement du signal Vol. 42; no. 1; pp. 277 - 289
• Main Authors: Nasreen, Shaik, Balasubramanian, Prakash, Badarudeen, Saleena, Rashid, Mamoon, Alshamrani, Sultan S., Wahab, Aisha Banu, AlGhamdi, Ahmed Saeed
• Format: Journal Article
• Language: English; French
• Published: Edmonton International Information and Engineering Technology Association (IIETA) 01.02.2025
• Subjects: Accuracy; Artificial neural networks; Brain; Brain cancer; Brain research; Classification; Data augmentation; Datasets; Deep learning; Feature selection; Genetic algorithms; Glioma; Image classification; Machine learning; Magnetic resonance imaging; Medical imaging; Real time; Spinal cord; Tumors; Wavelet transforms
• ISSN: 0765-0019; 1958-5608
• DOI: 10.18280/ts.420124

Brain tumors are abnormal growths of brain cells that can be benign (non-tumor) or malignant (tumor). These tumors can arise from different types of brain cells and occur in various brain regions. Timely detection is crucial for reducing the severity and improving prognosis. However, the traditional human examination suffers in early tumor detection due to the irregular patterns in MRI scans. Additionally, Machine learning and deep learning-based frameworks detect brain tumors more accurately than human analysis. This work introduces an efficient diagnostic approach with improved accuracy to classify the benign and malignant from MRI scans. This diagnostic approach consists of three levels. In the first level, the majority and minority samples are increased to train the framework with more subjects using ImageDataGenerator with real-time data augmentation. In the second level, a pre-trained Convolution Neural Network (CNN), namely the Xception framework, is utilized to learn comprehensive information about images. The hyperparameter tuning process improves the multi-class classification accuracy in the third level. The proposed framework classifies brain tumors into multiple such as glioma, meningioma, no tumor, and pituitary. The experimental dataset is obtained from the Kaggle repository to train the framework. The outcomes attained by the proposed framework deliberate higher accuracy compared with other CNN frameworks. The proposed framework proves its efficiency in the fine-grained classification of brain tumors with a validation accuracy of 99.87%. Thus, this framework may be employed in clinical services to diagnose brain MRI tumors.