Deep Transfer Learning-Based Xception Model for Robust Multi-Class Brain Tumor Diagnosis

Abstract

Correct brain tumor classification using magnetic resonance imaging (MRI) is essential for exact diagnosis, better patient outcomes, and treatment planning. This study proposes a robust multi-class brain tumor classification framework based on a fine-tuned Xception deep learning (DL) architecture, optimized through transfer learning and adaptive regularization. The pretrained Xception backbone, initialized on ImageNet, was customized with sequential dense layersFlatten, Dropout, Dense(128), Dropout, and Dense(4)-to categorize four types of tumors: pituitary, meningioma, glioma, and no tumor. The model achieved best convergence after Epoch 10, reaching a training accuracy of 99.09 %, training loss of 0.0261, precision of 0.9919, and recall of 0.9905. During the testing and validation phase, the model achieved a testing accuracy of 99.18 %, a loss of 0.0296, a precision of 0.9918, and a recall of 0.9918. The extremely small performance gap between training and validation confirms exceptional generalization capability and network stability. Overall, the proposed Xception-based framework proves outstanding diagnostic reliability, computational efficiency, and explainability, offering a powerful foundation for intelligent and automated brain-tumor detection in clinical MRI applications.