Leukemia White Blood Cell Classification Using DenseNet121 Embeddings and Ensemble Learning

Leukemia diagnosis through white blood cell (WBC) classification remains challenging, requiring expert pathologists and significant time investment. This study presents a hybrid approach for leukemia WBC classification using DenseNet121 embeddings combined with ensemble learning techniques. We utilize transfer learning with a pre-trained DenseNet121 model to extract 1024-dimensional feature embeddings from WBC images, which serve as input to various machine learning classifiers. Our methodology is evaluated on the comprehensive LeukemiaAttri dataset, which contains 14 different WBC types captured from two microscopes at three magnification levels. Experimental results demonstrate that tree-based ensemble methods, particularly CatBoost, XGBoost, and Multi-Layer Perceptron, achieve the best performance across different experimental settings. CatBoost achieves the highest accuracy of $\mathbf{5 6. 3 \%}$ on the H_100X_C2 configuration, while MLP reached $55.3 \%$ accuracy on H_100X_C1. Despite the dataset’s challenging nature due to image quality variations and class imbalance, our approach provides competitive results compared to previous YOLO-based methods. The study highlights the potential of embedding-based classification as an alternative to direct image-based deep learning models for leukemia diagnosis. It offers insights into classifier performance across various experimental conditions while maintaining computational efficiency.