DUET-Net: A Deformable CNN–Transformer with Calibrated Dual Uncertainty for Brain Tumor MRI Classification

Background: Magnetic resonance imaging (MRI) is a primary modality for brain tumor diagnosis and treatment planning, making accurate automated classification critically important. However, existing deep learning approaches face limitations in prediction reliability, generalization across imaging domains, and calibrated confidence estimation. To address these gaps, we introduce DUET-Net, a hybrid CNN–Transformer framework that integrates six complementary architectural innovations. Methods: First, a multiscale depthwise fusion (MSDF) module aggregates multi-receptive-field context from a single 2D MRI slice using parallel depthwise convolutions at three kernel sizes. Second, domain-adaptive batch normalization (DABN) with input-conditioned mixture weights maintains domain-specific statistics while sharing affine parameters, enabling sample-adaptive generalization across imaging sources. Third, an enhanced convolutional block attention module (CBAM) introduces deformable spatial offsets to adaptively attend to irregular tumor borders. Fourth, a deformable cross-scale attention (DCSA) module combines per-scale learnable spatial offsets with adaptive cross-scale weighting for geometry-aware multi-resolution fusion. Fifth, a hierarchical Vision Transformer encoder coupled with Monte Carlo dropout and a heteroscedastic aleatoric head enables dual uncertainty estimation grounded in the Kendall–Gal framework. Sixth, post-hoc temperature calibration and Grad-CAM++ provide trustworthy confidence scores and visual interpretability. Results: The proposed model achieves 99.51% accuracy on the Kaggle dataset, 95.27% on the Figshare dataset, and 99.89% on the Mendeley dataset with an expected calibration error (ECE) of 0.0142. With approximately 48 M parameters and 6.66 GFLOPs, DUET-Net sustains 52.41 FPS and is the only compared method to provide calibrated uncertainty estimation. Ablation, cross-dataset, and uncertainty-formulation comparison studies further validate the reliability of the framework. Conclusions: The potential of our proposed net as a promising approach for reliable computer-assisted brain tumor classification, although further external and prospective clinical validation is necessary before deployment.