Comparative performance analysis of binary variants of FOX optimization algorithm with half-quadratic ensemble...

Area of Science:

• Medical Diagnostics
• Machine Learning
• Endocrine System Oncology

Background:

• Thyroid cancer, though less common than other cancers, significantly impacts the endocrine system.
• Machine learning (ML) offers advanced tools for diagnosing thyroid abnormalities.
• Effective feature selection in ML is crucial for reducing data dimensionality, improving model performance, and enhancing interpretability.

Purpose of the Study:

• To investigate binary variants of FOX-optimization algorithms for effective feature selection in thyroid cancer detection.
• To evaluate the performance of these algorithms using vision transformer-based feature extraction and Naïve Bayes classification.
• To optimize the multi-objective task of maximizing accuracy, F2-score, AUC-ROC, and minimizing feature space size.

Main Methods:

• Employed eight transfer functions to create binary FOX-optimization algorithms.
• Utilized DeiT and Swin Transformer models for feature extraction from ultrasound and histopathological thyroid cancer images.
• Applied locally linear embedding for feature transformation, followed by binary FOX-optimization for feature selection with a Naïve Bayes classifier.

Main Results:

• The binary FOX-optimization algorithm with the V1 transfer function, combined with Swin Transformer and Naïve Bayes, demonstrated superior performance.
• Achieved 94.75% accuracy, 0.9848 AUC-ROC, and 0.9365 F2-Score on the ultrasound dataset, selecting only 5 features.
• For the histopathological dataset, the model reached 89.71% accuracy, 0.9329 AUC-ROC, and 0.8760 F2-Score, using 12 features.

Conclusions:

• The proposed framework effectively identifies thyroid cancer using a reduced feature space, comparable to existing research.
• The optimized binary FOX-optimization algorithm offers a highly accurate and efficient method for thyroid cancer diagnosis.
• This approach enhances diagnostic capabilities for endocrine system cancers through advanced machine learning techniques.