Hybrid Deep Learning Architecture Integrating ConvNeXtV2, Swin Transformer, and Convolutional Block Attention Module for Enhanced Multiclass Brain Tumor Classification in Magnetic Resonance Imaging
TL;DR
Overall, fusing convolutional and transformer features with attention refinement markedly enhances classification performance and achieves 95.45% accuracy versus RDXNet and ResNet50, with gains across all metrics.
Abstract
This study evaluates the performance of a hybrid deep learning framework against conventional approaches for multiclass brain tumor classification using MRI scans. The framework integrates ConvNeXtV2, the Swin Transformer, and Convolutional Block Attention Module attention mechanisms to capture both localized texture features and broader contextual patterns. RDXNet and ResNet50 serve as baseline models. Evaluation metrics include accuracy, precision, recall, macro F1-score, and area under the curve. The hybrid model achieves 95.45% accuracy versus RDXNet (94.00%) and ResNet50 (91.54%), with gains across all metrics. Gradient-weighted Class Activation Mapping visualizations confirm attention to tumor-relevant regions, supporting model interpretability. Overall, fusing convolutional and transformer features with attention refinement markedly enhances classification performance.