Optimized deep learning ensemble using Fast Osprey algorithm for accurate lymphoblastic leukemia detection

Area of Science:

• Hematology
• Computer Science
• Artificial Intelligence

Background:

• Acute Lymphoblastic Leukemia (ALL) is a life-threatening hematological malignancy requiring rapid diagnosis.
• Traditional diagnostic methods for ALL are labor-intensive and prone to inter-observer variability.
• Current deep learning models for ALL detection can overfit and lack interpretability.

Purpose of the Study:

• To develop a reliable and responsive computer-aided diagnostic (CAD) platform for ALL.
• To enhance the accuracy and reduce variability in ALL diagnostics.
• To improve patient outcomes through early and precise ALL detection.

Main Methods:

• An ensemble-based model combining EfficientNetB3, EfficientNetV2B3, and EfficientNetV2B1 was developed.
• The model was optimized using Fast Osprey Optimization (FOO) for dynamic ensemble weight assignment.
• Data augmentation was applied to an extensive dataset to address class imbalance and improve generalization.

Main Results:

• The FOO-Ensemble model achieved high performance: 97.76% accuracy, 98.13% precision, 97.71% recall, and 97.83% F1-score.
• The ensemble approach reduced inference time compared to individual models.
• The framework demonstrated robustness, scalability, and superior generalization capabilities.

Conclusions:

• Deep learning ensembles with bio-inspired optimization offer trustworthy ALL detection.
• Dynamic weighting mechanisms enhance stability and minimize overfitting risks.
• The FOO-Ensemble framework shows potential for clinical application, assisting hematopathologists in accurate ALL diagnosis.