In this paper, we introduce a deep learning pipeline for the enhancement of the clas-sification of White Blood Cells (WBCs) with Vision Transformers (ViTs) fortified with segmentatation masks and Kalman filtering. Our proposed pipeline revolves around a custom WBC Dataset class that combines labeled microscopy images with optional segmentation masks, augmenting the spacial context for improved predictions. The ViT model serves as the core point of high-level visual feature extraction and classification of the five main classes of WBCs. In order to further stabilize predictions and eliminate noise common in medical imaging we apply Kalman filtering as a post-processing step, smoothing the output feature trajectories over time. The data is subjected to standard preprocessing via torchvision transformations, and the evaluation procedure entails both visual input-mask pair visualizations and numerical metrics. Performance is assessed via classification re-ports and confusion matrices, which very clearly illustrate that the addition of Kalman smoothing yields superior consistency and accuracy over ViT alone. In general, tis hybrid methodology highlights the potential of merging attention-based models with temporal smoothing techniques to enhance the resilience of WBC classification, particularly in diagnostically problematic cases. Keywords: White blood cells, leukemia, Kalman filtering, classification, Feature Extraction, Image Analysis, CNN, Vision Transformer.