A novel pruning-enhanced hybrid approach for efficient and accurate brain tumor diagnosis

Abstract

Background: Accurate and efficient brain tumor diagnosis is a persistent challenge in medical imaging due to the complexity and variability of tumor structures. Deep learning has demonstrated strong potential; however, balancing diagnostic performance with computational efficiency remains a significant barrier to real-world deployment, particularly in resource-limited clinical settings. Proposed Method: This study introduces a novel hybrid framework combining a pruning-enhanced Effi-cientNetV2B3 model with metaheuristic-based feature selection and traditional machine learning classifiers. Multiple state-of-the-art CNN architectures were initially evaluated, and EfficientNetV2B3 was selected as the most effective model. A sparsity-driven threshold-based pruning strategy was then applied, reducing parameters by approximately 78% (from 12.9M to 2.86M) while retaining competitive classification accuracy. Feature vectors extracted from the pruned model were refined using Genetic Algorithm, Particle Swarm Optimization, and Lion Optimization Algorithm. Results: The pruned EfficientNetV2B3 model maintained high performance, achieving 97.70% +/- 0.11 accuracy. Among all evaluated classifiers, Random Forest combined with genetic Algorithm-selected features yielded the best results: 98.64% +/- 0.25 accuracy, 98.67% +/- 0.24 precision, and 98.66% +/- 0.26 F1-score. Additionally, visualization techniques such as t-SNE and Grad-CAM were used to confirm feature separability and enhance model interpretability. Conclusion/Impact: The proposed hybrid framework demonstrates that significant model compression can be achieved without compromising diagnostic accuracy. It offers a computationally efficient and interpretable solution suitable for deployment in real-time and low-resource clinical environments.