Automated bone marrow cell classification using ensemble learning: performance, generalization, and clinical...

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate bone marrow (BM) cell classification is crucial for diagnosing hematological disorders.
• Automated classification faces challenges like morphological overlap, class imbalance, and imaging artifacts.
• Deep learning models (CNNs) show promise but can lack generalization and robustness.

Purpose of the Study:

• To develop an ensemble-learning framework using MobileNetV3 and ResNet18 for enhanced BM cell classification.
• To improve feature extraction, classification performance, and interpretability while maintaining low computational cost.
• To evaluate the framework's accuracy, efficiency, and reliability in supporting AI-assisted hematological diagnostics.

Main Methods:

• An ensemble-learning framework combining MobileNetV3 and ResNet18 was developed.
• Four ensemble strategies (Soft Voting, Bagging, Boosting, Stacking) were evaluated on a large dataset (>420,000 images, 21 classes).
• Explainable AI (XAI) methods (Grad-CAM, LIME) were used for interpretability, with Decision Impact Ratio and Confidence Impact Ratio for reliability assessment.

Main Results:

• The Boosting ensemble strategy achieved the highest classification accuracy at 96%.
• External validation confirmed robust performance across independent datasets and varying imaging conditions.
• The ensemble model surpassed individual models in accuracy and interpretability stability, with XAI confirming focus on relevant morphological features.

Conclusions:

• The MobileNetV3-ResNet18 ensemble framework offers accurate, efficient, and interpretable BM cell classification.
• This approach can enhance diagnostic performance and explanation reliability for AI-assisted hematological diagnostics.
• The framework has the potential to reduce diagnostic time and interobserver variability.