Ensemble Deep Learning Approach for Early Detection and Subtype Classification of Blood Cancer Using Transfer Learning and Data Augmentation

Abstract

Blood cancer constitutes a significant worldwide health issue that needs precise diagnostic methods to enhance patient treatment results. The latest Deep Learning (DL) methods demonstrate promising results. However, the research studies face difficulties because of their unbalanced datasets and their insufficient feature extraction capabilities together with their missing effective multi-stage classification methods. The study aims to create a hybrid DL system that uses microscopic blood smear images to identify blood cancer at an early stage and determine its different subtypes. The proposed method combines data pre-processing and augmentation with deep feature extraction through Residual Network (ResNet50), followed by dimensionality reduction using Principal Component Analysis (PCA). The system uses Swin Transformer and Vision Transformer (ViT) together with their combined model to classify refined features, which results in better performance outcomes. The outcomes show that the proposed system accomplishes better classification accuracy because the ensemble model reached a 99.54% accuracy rate together with exceptional precision, recall, and F1-score metrics. The research results exhibit that the proposed method delivers an effective automated system for blood cancer diagnosis, which operates with high reliability. The research results demonstrate that advanced feature extraction methods together with dimensionality reduction techniques and hybrid transformer-based models, create better classification performance with increased system stability.