A Vision Transformer with a Self-Attention Mechanism for High-Accuracy Blood Cell Classification

Blood has several essential functions in the body, including immune defense against foreign elements and the transportation of oxygen, nutrients, and hormones. In this study, an IoT-based learning system was developed for classifying blood cell types, including basophils, erythroblasts, monocytes, myeloblasts, and segmented neutrophils. The study utilized a dataset of five blood cells, which was systematically split into three parts to ensure robust evaluations. A Convolutional Neural Network (CNN) model based on a Vision Transformer (ViT) combined with a Self-Attention Mechanism (SAM) was utilized to extract and learn discriminative features from cell images. To comprehensively evaluate and demonstrate execution, a suite of assessment metrics, classification reports, and Receiver Operating Characteristic (ROC) curves with the Area Under the Curve (AUC) was used in the training, validation, and testing phases. Overall, the proposed model achieved accuracies of 99%, 98%, and 98%, with balanced precision and recall, across the training, validation, and testing samples of five classes. Additionally, 5-fold cross-validation was conducted during the preparation and approval stages to improve generalizability and decrease overfitting. The model consistently demonstrated strong performance and maintained robust discriminative ability on the unseen test set. The proposed framework offers a promising solution to the limitations of conventional blood flow investigation by enabling digital, precise, and adaptable cell classification.