ENHANCED BRAIN TUMOR CLASSIFICATION VIA A FUSION ARCHITECTURE OF TRANSFER LEARNING AND DISCRETE COSINE TRANSFORM

Abstract

Brain tumors represent a major global health concern, where early and accurate diagnosis is critical for effective clinical intervention. However, reliable classification of brain tumor types from low-resolution MRI images remains challenging, while high-resolution images impose significant storage and computational demands. To address these issues, this study proposes a unified Deep Transfer Learning–based DCT (DTLT-DCT) framework that integrates Discrete Cosine Transform (DCT) with deep learning to simultaneously enhance classification accuracy, reduce image size, and improve peak signal-to-noise ratio (PSNR). Five pre-trained deep transfer learning architectures—VGG19, ResNet-152, DenseNet-201, EfficientNet-B0, and EfficientNet-B3—were evaluated after preprocessing and DCT-based compression. Experimental results demonstrate that the DCT-EfficientNet-B3 model achieved superior performance, with training, validation, and test accuracies of 99.94%, 99.72%, and 99.67%, respectively. Comparative analysis confirms that the proposed framework outperforms existing architectures across multiple evaluation metrics. These findings demonstrate the potential of integrating frequency-domain transformations with deep learning to improve diagnostic accuracy and efficiency in brain tumor classification, contributing to more effective decision-support systems in neuro-oncology.