NSGO: Convergence-Preserving Stochastic Gradient Oscillation for Brain Tumor Classification

Abstract

Conventional optimizers have been widely used for their fast convergence and robustness to gradient scaling during training via adaptive gradient-based techniques. However, these optimizers perform poorly at sharp minima, leading to reduced generalization and instability in brain tumor scan analysis on small and imbalanced datasets. In this work, we proposed a NAdam-Stochastic Gradient Oscillation (NSGO) optimizer, an upgrade to the NAdam optimizer that introduces controlled gradient oscillations while preserving standard convergence properties. We extensively compared NSGO with conventional optimizers across different conditions. We evaluated it on three benchmark datasets: Figshare (3-class), Sartaj (4-class), and Br35H (2-class), where the proposed NSGO optimizer consistently converged to more stable stationary points. The NSGO optimizer is evaluated on different performance metrics using 10-fold cross-validation, achieving accuracies of 96.16%, 94.75%, and 98.87% for Figshare (3-class), Sartaj (4-class), and Br35H (2-class), respectively. The proposed NSGO, with performance improvements of 3%, 4%, and 15%, respectively, outperforms the second-best NAdam optimizer. This provides an effective mechanism to improve optimization stability.