BRAIN TUMOR DETECTION FROM MRI IMAGES USING CNN, INCEPTIONV3, AND VISION TRANSFORMER

Abstract

Brain tumor detection from magnetic resonance imaging (MRI) is an important medical image-analysis task because early and reliable identification can support timely clinical assessment, treatment planning, and patient management. Brain tumors can exhibit considerable variations in size, shape, location, intensity, and texture, making their identification from MRI images a challenging task. Although manual interpretation by radiologists remains the standard diagnostic approach, the process can be time-consuming and may be influenced by image quality, anatomical complexity, workload, and inter-observer variability. Recent advances in artificial intelligence and deep learning have provided effective approaches for developing computer-aided diagnostic systems capable of automatically learning discriminative patterns from medical images. In this study, The proposed framework based on automated brain tumor detection using three different architectures: a custom Convolutional Neural Network (CNN), InceptionV3, and Vision Transformer (ViT). The proposed framework incorporates several stages, including image acquisition, data preprocessing, resizing, normalization, augmentation, dataset partitioning, model training, and performance evaluation. The CNN is employed as a baseline model for learning hierarchical spatial features directly from MRI images, while InceptionV3 utilizes transfer learning and multi-scale convolutional feature extraction to improve classification performance. The Vision Transformer adopts a patch-based representation and self-attention mechanism to capture both local and global contextual relationships within MRI images. The performance of the proposed models is evaluated using multiple statistical measures, including accuracy, precision, recall, F1-score, training and testing loss, and confusion matrices, providing a comprehensive assessment of their classification capabilities. In the illustrative experimental configuration presented in this paper, the CNN achieves a test accuracy of 96.84%, while InceptionV3 achieves 98.21% and the Vision Transformer achieves 99.12%. The corresponding macro-averaged precision, recall, and F1-scores remain above 96% for all three models, demonstrating strong classification performance across the evaluated approaches. Comparative analysis indicates that InceptionV3 provides an improvement over the conventional CNN through pretrained feature representations and multi-scale feature extraction, whereas the Vision Transformer achieves the highest performance by exploiting self-attention to learn broader relationships among different image regions. These findings demonstrate the potential of advanced deep learning architectures for automated brain tumor detection and computer-aided MRI analysis. The proposed framework may assist healthcare professionals by providing rapid and consistent preliminary predictions.

Keywords: Brain tumor detection; MRI; deep learning; CNN; InceptionV3; Vision Transformer; ViT; medical image classification; transfer learning; computer-aided diagnosis