Morfoloji Tabanli Hibrit Hiyerarşik Periferik Kan Hücresi Siniflandirmasi

The accurate differentiation of morphologically similar sub-types is a fundamental challenge in the automated classification of peripheral blood cells. In this study, a feature embedding-based hybrid hierarchical deep learning architecture is proposed for the precise analysis of these classes. The proposed system consists of a primary DenseNet121 model followed by specific expert models that examine nuclear and cytoplasmic de-tails across morphological variations. By integrating the extracted features into a machine learning layer, an overall accuracy of 98.07% is achieved. Notably, in the Band Neutrophil class where standard models often underperform, the F1-score is increased from 0.339 to 0.886, and in the Myelocyte class, it rises from 0.138 to 0.737. The results demonstrate that hierarchical morphological analysis provides high performance in the differential diagnosis of complex cell groups.