Home / Projects / NSGA-II vs Optuna
Under Review IEEE · 2025

A Comparative Study of NSGA-II and Optuna for Brain Tumor Classification

Comparing evolutionary and Bayesian multi-objective optimization for energy-efficient brain tumor classification.

Overview

This research investigates the use of NSGA-II and Optuna (Tree-structured Parzen Estimator) for multi-objective hyperparameter optimization of convolutional neural networks for brain tumor classification. Unlike conventional optimization approaches that prioritize predictive accuracy alone, this work jointly optimizes validation accuracy, Macro F1-score, AUC-ROC, model size, training time, and inference time to develop computationally efficient deep learning models suitable for deployment in resource-constrained medical environments.

Research Objectives

  • Compare NSGA-II and Optuna for multi-objective hyperparameter optimization.
  • Balance predictive performance with computational efficiency.
  • Evaluate trade-offs between accuracy, model size, training time, and inference speed.
  • Identify the most suitable optimization strategy for different deployment scenarios.

Methodology

  • Brain MRI dataset containing 7,023 images across four tumor classes.
  • TensorFlow-based CNN optimized using NSGA-II (DEAP) and Optuna.
  • Optimization objectives included Validation Accuracy, Macro F1-score, AUC-ROC, Training Time, Inference Time, and Model Size.
  • Performance evaluated using Pareto fronts and spread metrics.

Results

  • NSGA-II achieved 98% validation accuracy while producing a diverse Pareto-optimal solution space.
  • Optuna converged in approximately 2 hours, compared to roughly 2 days for NSGA-II.
  • Optuna generated significantly smaller models (1.5 MB) while NSGA-II produced larger but more diverse solutions (12.3 MB).
  • Both approaches achieved similar predictive performance, demonstrating the trade-off between optimization speed and solution diversity.

Conclusion

NSGA-II is well suited for discovering diverse trade-offs between competing objectives, making it appropriate for comprehensive multi-objective optimization studies. Optuna offers substantially faster optimization and lightweight models, making it ideal for rapid experimentation and deployment where computational resources are limited. The study highlights that the choice of optimizer depends on whether solution diversity or optimization efficiency is the primary objective.

Gallery

Presentation & Videos