D2EF-Net: A Deep Dual-Encoder Fusion-Based Unified Network for Brain Tumor Classification from Undersampled MRI

Magnetic resonance imaging (MRI) offers superior diagnostic quality but suffers from prolonged acquisition times, leading to patient discomfort and motion artifacts. The challenge of undersampled MRI data adversely impacts brain tumor classification accuracy. To address this, we propose D2EF-Net, a unified framework for brain tumor classification from undersampled MRI data. The model integrates MRI reconstruction and classification into a joint learning framework, preserving key diagnostic features while improving accuracy. D2EF-Net introduces three novel modules: adaptive multiscale convolution (AMC) for efficient feature extraction, residual depthwise convolution (RDC) for reduced complexity, and attention-enhanced hybrid transformer (AHT) for comprehensive feature representation. Extensive experiments on five datasets (DS-1 to DS-5) demonstrate that D2EF-Net significantly outperforms existing methods in tumor classification accuracy. Notably, it achieved average improvements of 4.66%, 4.61 %, 14.94%, 10.01%, 28.53%, 10.07%, 10.82%, 4.60%, 26.01 %, and 7.97% over baseline models. Additionally, D2EF-Net excels in fully-sampled data scenarios, further showcasing the flexibility of its joint learning mechanism. In conclusion, D2EF-Net offers a robust solution for accelerating MRI acquisition while maintaining high diagnostic accuracy, with potential applications in clinical practice.