WHITE BLOOD CELL CLASSIFICATION USING SERIAL DEEP FEATURE FUSION AND GENETIC ALGORITHM FEATURE SELECTION

Abstract

This study reports a white blood cell (WBC) classification framework that combines deep feature extraction, serial feature fusion, genetic algorithm (GA)-based feature selection, and support vector machine (SVM) classification. A public blood-cell dataset comprising 12,500 enhanced JPEG images with CSV cell-type labels for Eosinophil, Lymphocyte, Monocyte, and Neutrophil was used. AlexNet, ResNet-101, and DarkNet-53 extracted 4096, 1024, and 1024 features, respectively. Serial concatenation produced a 6144-feature representation, which was reduced using a GA with Roulette Wheel Selection and a crossover rate of 0.5. Six SVM variants were evaluated using the reported five-fold validation procedure. Three selected-feature configurations were examined: 500, 2500, and 3130 features. The strongest detailed result was reported for Cubic SVM with 3130 features: 98.9% accuracy, 0.99 precision, 0.9925 recall, and 0.99 F1-score. Under the reported configuration, the 3130-feature setting produced the highest performance among the reported experiment rows.