Chaotic Harris Hawks Optimization with Quasi-Reflection-Based Learning: An Application to Enhance CNN Design

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Intelligence

Background:

• Accurate brain tumor classification is crucial for effective treatment planning.
• Convolutional Neural Networks (CNNs) show promise in medical image analysis but require optimized architectures.
• Swarm intelligence algorithms offer potential for improving CNN design.

Purpose of the Study:

• To propose an improved Harris Hawks optimization (HHO) algorithm for evolving CNN architectures.
• To apply the optimized CNN for classifying brain tumor grades using MRI.
• To enhance the exploration and exploitation capabilities of the HHO algorithm.

Main Methods:

• Developed an enhanced HHO algorithm incorporating chaotic population initialization, local search, and quasi-reflection-based learning.
• Evaluated the enhanced HHO on CEC2019 benchmarks against basic HHO and state-of-the-art methods.
• Applied the HHO-evolved CNN to classify brain tumors using two MRI datasets (IXI, Cancer Imaging Archive, and T1-weighted images) with data augmentation.

Main Results:

• The enhanced HHO algorithm demonstrated superior performance on benchmark functions.
• The HHO-evolved CNN achieved over 95% accuracy in classifying healthy brains and various tumor grades (I-IV) and types (Glioma, Meningioma, Pituitary).
• The proposed swarm intelligence-driven CNN approach outperformed existing methods in empirical validation.

Conclusions:

• The novel, enhanced Harris Hawks optimization algorithm effectively evolves CNN architectures for brain tumor classification.
• The developed approach shows significant potential for assisting clinicians in early brain tumor detection and diagnosis.
• The method offers a robust, high-accuracy solution for automated analysis of magnetic resonance imaging data.