Stacking ensemble techniques for automated peripheral blood cell classification using Inception v3 features

Abstract

Robust distinction of blood cells is crucial in clinical evaluation. Manual examination is slow and exposed to errors. This work investigates using machine learning (ML) techniques for automated classification of eight categories of peripheral blood cell types from multi-color images. The Inception v3 network was used to extract features, a split of 66%/34% were used to evaluate the model along with 20-fold cross-validation. To reduce computational complexity, principal component analysis (PCA) was used to reduce the 2048-dimensional feature vectors to 100 components. Among all classifiers used, the highest performance without using PCA was achieved using the support vector machine (SVM) with an accuracy equal to 93.4% and an area under the curve (AUC) of 0.996. Using PCA, affected monocytes and immature granulocytes most due to the slight reduction in the accuracy and AUC which became 90.1% and to 0.991 respectively. Results were further enhanced when a stacked ensemble of neural network (NN), logistic regression (LR), and SVM were used, achieving an accuracy of 95.2% and an AUC of 0.998. The obtained findings confirmed the effectiveness of using stacked ensembles in providing a robust, high accuracy framework for automated blood cell classification, while PCA efficiently reduced dimensions with minimal performance loss.