Evaluating the classification performance of a ViT, DeiT and Swin-based hybrid deep learning model for white blood cell images

White blood cells (WBCs) are essential components of the immune system, playing a critical role in defending the body against pathogens and diseases. Accurate classification of WBCs is particularly important for the early diagnosis of hematological disorders and for guiding effective treatment planning. However, manual classification of WBC images is both time-consuming and prone to inter-observer variability. To address this challenge, deep learning-based image classification approaches have become increasingly prevalent. In recent years, the use of transformer architectures, capable of extracting more effective features through attention mechanisms, has gained popularity in this domain. In this study, we developed a novel ensemble model that integrates the strengths of multiple transformer architectures to enhance WBC classification performance. The proposed model was trained and evaluated on three publicly available datasets: Raabin-WBC, Leukocyte Images for Segmentation and Classification (LISC), and Peripheral Blood Cells (PBC). To mitigate class imbalance, synthetic images were generated using a Deep Convolutional Generative Adversarial Network (DCGAN) solely for the Raabin-WBC dataset. Additionally, all dataset images were preprocessed using Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance contrast and highlight relevant cellular structures. The ensemble model combines three state-of-the-art transformer-based architectures: Vision Transformer (ViT), Data-efficient Image Transformer (DeiT), and Swin Transformer. Its classification performance was benchmarked against several widely-used convolutional neural networks (CNNs), including ResNet50, InceptionV3, VGG19, and DenseNet121. Furthermore, the classification results were interpreted visually using the Score-CAM technique to highlight the regions that contributed most to the model's decisions. Our model demonstrated superior performance across all datasets. It achieved classification accuracies of 99.83% on the original Raabin dataset, 99.60% on the balanced Raabin dataset, 99.81% on the LISC dataset, and 99.84% on the PBC dataset—surpassing previously reported results in the literature. These findings suggest that the proposed transformer-based ensemble model offers significant improvements in both accuracy and generalization over traditional CNN based approaches.