Enhancing Blood Cancer Detection and Classification Using Deep Ensemble Learning Approaches

Abstract

Blood cancers such as leukaemia pose a severe threat to human health, with hundreds of thousands of new cases each year worldwide. Early and accurate diagnosis is critical for improving patient outcomes. However, manual microscopic examination of blood smears is time-consuming and prone to observer variability and error. In recent years, deep learning particularly convolutional neural networks (CNNs) has achieved remarkable success in medical image analysis including automated detection of leukaemia. Yet distinguishing between different leukaemia subtypes in images can be challenging due to subtle morphological differences and limited data. In this paper, we propose a deep ensemble learning approach to enhance the accuracy of blood cancer detection and subtype classification. We develop an ensemble of CNN models that combines the strengths of multiple architectures to improve generalization. Experiments are conducted on a public blood cell image dataset of 10,000 single-cell images augmented with healthy cell images for binary classification. The proposed ensemble achieves superior performance compared to individual models, with an overall classification accuracy of about 98.5% on test images a significant improvement over single CNNs (which achieved ~95%–96%). The results demonstrate that deep ensembles can provide more robust and accurate diagnosis of leukaemia from blood smear images. This approach could assist pathologists by providing reliable second opinions and aiding early detection. We also discuss the implications of model interpretability and the need for diverse training data. Overall, the ensemble deep learning strategy shows great promise in improving computer-aided diagnosis of blood cancers, potentially leading to better clinical decision-making and patient outcomes.