A Hybrid VGG16-ViT Approach With Image Processing Techniques for Improved White Blood Cell Classification and Disease Diagnosis: A Retrospective Study.

Background and Aims: White Blood Cells (WBCs) are essential for immune defense against infections. Automated WBC identification from microscopic images aids in diagnosing diseases like leukemia and AIDS. However, the complexity of WBC morphology due to varying maturation stages and staining techniques complicates classification. This study aims to enhance WBC detection and classification using a hybrid VGG16-Vision Transformer (VGG16-ViT) model. Methods: To enhance the efficiency of the classification process, preprocessing techniques such as data normalization, categorical variable encoding, feature extraction, and data augmentation were employed in conjunction with the proposed model before the training phase. The VGG16-ViT model was trained and evaluated on two datasets to measure its performance. Results: The overall success rate for classifying WBCs was 98.12% for Dataset 1% and 99.60% for Dataset 2. The measured average precision, recall, and F1-score values were 98.59%, 98.23%, and 98.35% for Dataset 1; similarly, 98.95%, 99.98%, and 99.48% for Dataset 2. The experimental results indicated that the classification success was strengthened when the proposed model was combined with specific preprocessing procedures, outperforming existing research. Conclusion: The hybrid VGG16-ViT model, combined with effective preprocessing techniques, significantly improved the detection and classification of WBCs. Additionally, the training approach of the proposed model is less time-consuming than existing transfer learning models, making it a valuable tool for assisting medical professionals in diagnosing diseases related to WBCs.