MSPO: A machine learning hyperparameter optimization method for enhanced breast cancer image classification

Area of Science:

• Medical imaging
• Artificial intelligence
• Computational biology

Background:

• Breast cancer is a major global health threat requiring early diagnosis.
• Deep learning shows promise in breast cancer image classification but faces hyperparameter optimization challenges.
• Conventional optimization methods often suffer from limited effectiveness and premature convergence.

Purpose of the Study:

• To propose and evaluate a novel Multi-Strategy Parrot Optimizer (MSPO) for breast cancer image classification.
• To enhance the performance of deep learning models in medical image analysis.
• To address the limitations of existing hyperparameter optimization techniques.

Main Methods:

• Developed MSPO by integrating Sobol sequence initialization, nonlinear decreasing inertia weight, and a chaotic parameter into the original Parrot Optimizer.
• Validated MSPO's performance on CEC 2022 benchmark functions and conducted an ablation study on its variants.
• Combined MSPO with the ResNet18 model for breast cancer image classification on the BreaKHis dataset.

Main Results:

• MSPO demonstrated superior optimization precision and convergence rate compared to leading algorithms on benchmark functions.
• The MSPO-optimized ResNet18 model significantly outperformed non-optimized versions and alternative optimization algorithms on the BreaKHis dataset.
• The ablation study confirmed the effectiveness of individual strategies within MSPO.

Conclusions:

• MSPO offers enhanced global exploration and convergence steadiness for optimization tasks.
• The proposed MSPO shows significant potential and practical value for medical image classification, particularly for breast cancer.
• Optimizing deep learning hyperparameters with MSPO improves diagnostic accuracy and classification performance in breast cancer detection.