Optimizing blood cell classification: evaluating feature dimensionality and validation strategies

Abstract

Manual blood cell classification is time consuming and may lead to inconsistent results. This study aims to assist pathologists in diagnosing hematological disorders using machine learning (ML) techniques for automated classification of blood cells in multi-color test images, distinguishing red blood cells (RBCs) and white blood cells (WBCs). Features were extracted using the InceptionV3 network, and several ML models were evaluated for classifying blood cells into eight categories. Two validation strategies: a 66%–34% train–test split and 20-fold cross-validation were applied. The effect of dimensionality reduction through principal component analysis (PCA) was also examined, reducing the feature space from 2,048 to 100 components. Among all models, support vector machine (SVM) achieved highest performance, with 93.4% accuracy and an area under the curve (AUC) of 0.996 without PCA, and 90.1% accuracy with an AUC of 0.991 after PCA. Although PCA slightly reduced accuracy, it improved computational efficiency. Overall, SVM provided the most accurate, stable, and generalizable classification results for automated blood cell analysis.