ALL classification using neural ensemble and memetic deep feature optimization

Area of Science:

• Hematology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Acute lymphoblastic leukemia (ALL) is a critical blood disorder requiring precise detection for effective treatment.
• Deep convolutional neural networks (CNNs) show potential in digital pathology but struggle with subtle leukemia subtype differences.
• Accurate classification of ALL subtypes is crucial for prognosis and personalized treatment strategies.

Purpose of the Study:

• To develop an improved pipeline for the binary detection and subtype classification of ALL using blood smear images.
• To enhance the accuracy and efficiency of ALL diagnosis through advanced deep learning and optimization techniques.
• To address the challenges posed by subtle morphological variations among leukemia subtypes.

Main Methods:

• A customized 88-layer deep CNN was developed and trained using transfer learning with GoogleNet CNN for feature ensemble creation.
• Feature selection was modeled as a combinatorial optimization problem, employing a memetic binary whale optimization algorithm with Differential Evolution.
• The proposed approach was validated on standard public datasets of peripheral blood smear images.

Main Results:

• Achieved an overall best average accuracy of 99.15% for binary ALL classification, with a significant 85% reduction in feature vector size.
• Demonstrated high performance metrics for binary classification: 99% precision and 98.8% sensitivity.
• Attained 98.69% accuracy for B-ALL subtype classification, with 98.7% precision and 99.57% specificity.

Conclusions:

• The proposed deep learning pipeline significantly improves the accuracy and efficiency of acute lymphoblastic leukemia detection and subtype classification.
• The integration of CNNs with advanced optimization algorithms offers a powerful tool for digital pathology in hematological diagnostics.
• This methodology outperforms existing studies, highlighting its potential for clinical application in diagnosing ALL.