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

Bone marrow (BM) cell classification is essential for diagnosing a wide range of hematological disorders. However, automated classification remains challenging due to morphological overlap among cell types, class imbalance, and imaging artifacts. Although deep learning models, particularly convolutional neural networks (CNNs), have shown strong potential in medical image analysis, individual models may suffer from limited generalization and robustness.

To address these limitations, we developed an ensemble-learning framework based on MobileNetV3 and ResNet18 to enhance feature extraction and classification performance while maintaining low computational cost. We evaluated four ensemble strategies: Soft Voting, Bagging, Boosting, and Stacking using a largescale dataset comprising more than 420,000 images across 21 bone marrow cell classes. External validation was performed using independent datasets to assess model generalizability under different imaging conditions. To improve interpretability, explainable AI methods, including Grad-CAM, Grad-CAM++, and LIME, were applied to visualize discriminative image regions. In addition, Decision Impact Ratio and Confidence Impact Ratio were used to quantify the reliability of the generated explanations.

Among the evaluated ensemble strategies, Boosting achieved the highest classification accuracy of 96%. External validation confirmed that the proposed model maintained robust performance across independent datasets and varying imaging conditions. The ensemble model outperformed individual models in both classification accuracy and interpretability stability. XAI analysis further demonstrated that the model focused on relevant morphological features, supporting the clinical plausibility of its predictions.

The proposed MobileNetV3-ResNet18 ensemble framework provides accurate, computationally efficient, and interpretable bone marrow cell classification. By improving diagnostic performance and explanation reliability, this approach has strong potential to support AI-assisted hematological diagnostics, reduce diagnostic time, and minimize interobserver variability.