Cumulative Learning-Based Innovative Optimization Framework for Brain Tumor Classification

Abstract

The accurate classification of brain tumors using MRI scans has been a crucial part of medical imaging analysis. However, class imbalance in the data, overfitting, and unstable convergence of the optimizer are significant concerns in a conventional deep learning framework. The proposed framework introduces an innovative optimization approach that combines Nadam and Particle Swarm Optimization (PSO), achieving a balance between adaptive and robust convergence. This approach leverages the global search ability of PSO across the entire data space while employing a cumulative learning strategy to efficiently train the ResNet-50 model. The proposed framework is trained using a cumulative learning framework, which improves generalization by gradually improving learned representations across a series of learning phases. The proposed classification framework is evaluated using the Figshare, Br35H, and Sartaj datasets, yielding accuracies of 97.44%, 99.33%, and 99.70%, respectively.