Hybrid CNN-Based Transfer Learning Enhances Brain Tumor Classification on MRI Images
Rizal Dwi Prayogo, Nur Hamid, Hidetaka Nambo
Abstract
Brain tumors are among the deadliest diseases worldwide and require early and accurate diagnosis via Magnetic Resonance Image (MRI). Deep learning techniques, particularly convolutional neural networks (CNNs), have been widely adopted for analyzing brain MRI images in tumor classification tasks. However, individual models of conventional CNNs often struggle with tumor diversity due to their restricted receptive fields and homogeneous feature extraction, leading to suboptimal diagnostic precision.We propose a hybrid transfer learning framework based on CNN architectures for enhanced multiclass brain tumor classification. A publicly accessible brain MRI dataset from Kaggle is employed in this study. The dataset, constructed from Figshare, SARTAJ, and Br35H, contains 7023 original images categorized into four classes: glioma, meningioma, pituitary, and no tumor.We introduce a preprocessing pipeline involving MRI cropping and affine-based augmentation to balance a strategically modified version of our dataset, which includes intentional class imbalance. Leveraging transfer learning, we fine-tune lightweight pre-trained models such as ResNet50V2, MobileNetV2, DenseNet121, EfficientNetV2S, and NASNetMobile and adopt a hybrid feature extraction strategy that integrates multilevel feature maps. Experimental evaluations reveal that our hybrid model (ResNet50V2 + MobileNetV2 + DenseNet121) achieves superior performance, attaining 98.75% accuracy, 98.76% precision, 98.75% recall, and 98.75% F1-score, outperforming individual models and state-of-the-art methods. The results highlight that integrating diverse features from multiple CNNs enhances classification robustness by capturing complementary tumor characteristics. In conclusion, our study advances medical imaging diagnostics by demonstrating the efficacy of hybrid transfer learning in overcoming data imbalance and model limitations, offering a reliable tool for clinical decision-making.