A feature-fusion ensemble deep learning framework for accurate and automated leukaemia diagnosis

Leukaemia is a haematological malignancy that requires early and accurate diagnosis for effective treatment. Conventional manual techniques for examining smear samples are time-consuming and prone to human error. This study introduces an ensemble architecture for leukaemia diagnosis. The proposed design, a modified Convolutional Neural Network (MCNN), integrates hierarchical deep features from ResNet50 with manually constructed descriptors, specifically Local Directional Patterns (LDP) and Histogram of Oriented Gradients (HOG), to enhance feature extraction and overcome inherent CNN limitations such as overfitting. The combination of baseline CNN, ResNet50, and MCNN yielded the best diagnostic performance, achieving 96.50% accuracy, 96.30% precision, 96.70% recall, and a 96.50% F1-score over 50 training epochs, achieving better performance than individual MCNN (85.47% accuracy) and ResNet50 (88.50% accuracy) baselines. The ensemble of deep hierarchical features and handcrafted descriptors within the CNN framework simplifies the differentiation between leukaemic cells across various datasets. This ensemble framework is a generalised model that can significantly support clinical decision support systems, particularly in resource-limited healthcare environments.