Enhanced multiclass blood cell classification using contourlet transform and metaheuristic-optimized deep features with clustering-based decision making

Accurate classification of white blood cells (WBCs) is critical for the diagnosis of hematological disorders. This paper presents a novel hybrid framework that integrates multi-scale feature extraction, deep learning, metaheuristic optimization, and clustering-based decision making for robust multiclass WBC classification. The proposed method first employs the contourlet transform to decompose microscopic blood smear images, effectively capturing intricate contour and directional edge information in the frequency domain. These features are then sequentially processed by a recurrent neural network (RNN) to model hierarchical dependencies. To enhance discriminative power and reduce computational complexity, the African Vulture Optimization Algorithm (AVOA) is leveraged for optimal feature selection. Finally, a fuzzy clustering-based decision strategy is introduced to refine the classification of five WBC subtypes: lymphocytes, monocytes, eosinophils, basophils, and neutrophils. The framework emphasizes not only high accuracy but also operational efficiency, addressing key requirements for clinical deployment. Experimental evaluation on the Jiangxi Tecom dataset demonstrates superior performance over baseline models, with significant improvements in precision, recall, and F1-score across most cell types. Despite the inherent class imbalance for basophils, the model maintains viable performance, with augmentation techniques identified for future enhancement. The study's primary contribution lies in the unified integration of Contourlet-based feature extraction with deep sequential learning and metaheuristic-driven feature selection, offering a promising computer-aided diagnostic tool for automated hematological analysis.