TAC-Net: A Trustworthy Attention Cross-Fusion Network for Noise Robust and Explainable Multi-Class Brain Tumor Classification from MRI

Abstract

Accurate, robust, and interpretable classification of brain tumors using magnetic resonance imaging (MRI) continues to pose challenges, even with the progress made in deep learning technologies.This paper proposes TAC-Net, a Trustworthy Attention Crossfusion Network that combines an EfficientNet-B0 convolutional branch with a Swin-Tiny transformer branch through a Cross-Modal Gated Attention Fusion (CMGAF) module that learns a spatially adaptive gate balancing local and global evidence. The network is trained with a Noise-Adaptive Curriculum Augmentation strategy exposing it to Gaussian noise, motion blur, salt-andpepper corruption, and contrast shifts, and incorporates Monte Carlo dropout for predictive uncertainty and Grad-CAM++ for visual explanation. TAC-Net achieves 99.12% accuracy and 0.998 (macro-averaged) AUC on a benchmark of 7023 images derived from Figshare, outperforming ResNet50 and Swin-Tiny-only and feature-fusion baselines, and maintains 94.6% accuracy even with severe salt-and-pepper noise. Ablation studies show that the contributions from fusion, augmentation, and uncertainty calibration are complementary.