Attention guided hybrid fusion of MobileNetV3 DenseNet121 and DINOv2 for five class white blood cell classification

Abstract Accurate white blood cell classification is an important component of hematological image analysis, yet robust automation remains challenging because microscopic smear images often exhibit substantial morphological similarity across subtypes, staining variation, acquisition heterogeneity, and sensitivity to surrounding background content. Existing single-backbone approaches may not fully capture the complementary local texture, cellular morphology, and global contextual information required for reliable recognition across heterogeneous white blood cell images, while the effects of preprocessing choices such as cropping and image standardization are often insufficiently examined. To address these issues, this study proposes a three-branch hybrid fusion architecture for five-class white blood cell classification on a combined PBC-Normal and Raabin-WBC dataset comprising 27,606 images, including 3337 basophils, 6158 eosinophils, 4685 lymphocytes, 4,108 monocytes, and 9318 neutrophils. The framework integrates MobileNetV3, DenseNet121, and pretrained DINOv2 feature branches to capture lightweight convolutional cues, densely reused local representations, and transformer-based global semantic features, respectively. Branch-wise 512-dimensional embeddings are combined through feature stacking, layer normalization, multi-head attention, and a residual-style classification head for final five-class prediction. Performance was examined under four image-preparation settings, namely not cropped and unprocessed, not cropped and processed, cropped and unprocessed, and cropped and processed, together with dataset-specific analyses on PBC-Normal and Raabin-WBC. Across these comparative experiments, the proposed model achieved the strongest overall performance, with particularly high results on the PBC-Normal dataset and consistently favorable behavior across the combined-dataset evaluation settings, while processed inputs generally improved convergence and classification quality relative to unprocessed counterparts. These findings indicate that attention-guided CNN-transformer fusion, combined with controlled preprocessing evaluation, provides an effective and robust framework for automated five-class white blood cell classification across heterogeneous microscopic images.