CAE-BrainNet: a statistically validated class-adaptive attention ensemble model for explainable brain tumor classification from MRI.
Authors
Affiliations (4)
Affiliations (4)
- Multidisciplinary Action Research (MARS) Lab, Department of Computer Science and Engineering, Daffodil International University, Birulia, Dhaka, 1216, Bangladesh.
- Multidisciplinary Action Research (MARS) Lab, Department of Computer Science and Engineering, Daffodil International University, Birulia, Dhaka, 1216, Bangladesh. [email protected].
- Department of Computer Science and IT, La Trobe University, Melbourne, VIC, Australia. [email protected].
- Peter Faber Business School, Australian Catholic University, Brisbane, QLD, Australia.
Abstract
Deep learning has further accelerated progress in automated Magnetic Resonance Imaging (MRI) based classification of brain tumors, whereas prior studies have often reflected critical limitations such as narrow classification scope, insufficient Explainable Artificial Intelligence, weak statistical validation, and incomplete deployment metrics. These ultimately impede clinical trust and practical deployment. This study presents the CAE-BrainNet, a statistically validated Class-Adaptive Attention Ensemble model, which integrates representations of EfficientNetV2-M, DenseNet201, and ConvNeXt-Base dynamically, with complementary inductive biases based on different tumor morphology. Unlike uniform ensembles, a learned class-adaptive attention mechanism, parameterised by a [Formula: see text] weight matrix with a trainable temperature scalar, adaptively leverages model-specific morphological strengths, serving as the primary performance driver confirmed by ablation analysis. A hybrid optimization strategy based on grid search refinement via gradient-based optimization is also employed to estimate optimal hyperparameters. In addition, McNemar's and Cochran's Q tests were conducted, proving that the accuracy gains are statistically significant and not stochastic in nature. Experimental results show that the proposed CAE-BrainNet achieves 99.39% accuracy on a challenging four-class benchmark, with clinically critical 99.89% specificity in the identification of healthy tissue. To overcome the limitations of narrow scope and generalizability, zero-shot transfer assessment on the independent BRISC2025 dataset yields 98.10% accuracy, statistically validated against all base models, confirming cross-dataset robustness without retraining. Quantitative explainability via Grad-CAM, AWCLF, and IoU spatial agreement, combined with formal hypothesis testing and publicly accessible deployment at https://huggingface.co/spaces/Meraj-21/Brain-Tumor-Classification-Using-Grad-CAM , makes CAE-BrainNet a rigorous, transparent, and clinically deployable benchmark for brain tumor classification.