Hybrid Feature-Learning-Based PSO-PCA Feature Engineering Approach for Blood Cancer Classification

Area of Science:

• Computational Biology and Bioinformatics
• Medical Imaging and Diagnostics
• Machine Learning and Artificial Intelligence

Background:

• Acute Lymphoblastic Leukemia (ALL) is a critical blood cancer requiring early detection for effective treatment.
• Current manual screening of blood smear images for ALL is labor-intensive and prone to errors.
• Deep learning computer vision systems show promise for ALL detection but can suffer from feature redundancy and dimensionality issues.

Purpose of the Study:

• To develop an advanced feature engineering approach for enhanced ALL detection in blood peripheral images (BPIs).
• To integrate Principal Component Analysis (PCA) and Particle Swarm Optimization (PSO) for optimal feature selection.
• To improve the classification accuracy and efficiency of ALL diagnosis systems.

Main Methods:

• Image features were extracted using GoogleNet (a pre-trained Convolutional Neural Network - CNN).
• Principal Component Analysis (PCA) was applied to retain 95% data variability, and Particle Swarm Optimization (PSO) was used for optimal feature searching.
• A hybrid feature set combining PCA and PSO outputs was created to train Bayesian-optimized Support Vector Machine (SVM) and Subspace Discriminant Ensemble Learning (SDEL) classifiers.

Main Results:

• The proposed hybrid feature set significantly improved classification performance compared to individual PCA, PSO, or raw extracted features.
• The Bayesian-optimized SVM classifier trained with the hybrid PCA-PSO feature set achieved a high classification accuracy of 97.4%.
• The developed feature engineering approach demonstrates competitive performance against current state-of-the-art methods for ALL multi-class classification.

Conclusions:

• The hybrid PCA-PSO feature engineering approach effectively addresses dimensionality challenges in deep learning for ALL detection.
• This novel method enhances the accuracy and efficiency of machine learning classifiers in diagnosing ALL from BPIs.
• The study highlights the potential of integrating dimensionality reduction and evolutionary computation for robust medical image analysis.