ASO‑Optimized EfficientNet‑B2 with CBAM for Brain Tumor MRI Classification

Abstract

Magnetic resonance imaging (MRI) plays a critical role in brain tumor diagnosis; however, accurate classification remains a challenging task in medical image analysis. Gradient-based optimizers form the foundation of deep learning models; however, they are susceptible to suboptimal local minima and hyperparameter sensitivity. This paper proposes a hybrid model in which the CBAM attention module, the classifier head, and the final two convolutional blocks of EfficientNet-B2 are optimized using Atom Search Optimization (ASO), while the remaining backbone layers are retained as fixed feature extractors. ASO generates a population of candidate solutions by simulating atomic motion using Lennard-Jones and bond-length potentials, without relying on gradient information, and is used to search the high-dimensional parameter space. Experiments on the Figshare brain tumor dataset (3,064 T1-weighted MRI images of glioma, meningioma, and pituitary tumors) demonstrate that the ASO-optimized EfficientNet-B2+CBAM model has an accuracy of 99.35% and F1-score of 99.38%, outperforming Adam (98.76%) and SGD (98.12%). Statistically significant improvements (p < 0.01) are observed, and convergence analysis indicates ASO is less susceptible to plateaus of loss, and the variance is reduced by 44.7% across runs. The effectiveness of ASO is further evaluated through ablation studies, ROC analysis, and Grad-CAM visualizations, demonstrating its potential as an alternative optimizer for brain tumor MRI classification. Further validation on multi-center clinical datasets is required to assess generalizability and clinical applicability.