A new deep learning framework that fuses three different kinds of visual intelligence has reported some of the strongest results yet in the automated classification of white blood cells, reaching an overall accuracy of 99.3% on one benchmark dataset and 97.9% on a larger, more heterogeneous combined collection. The study, published as an open-access research article in Discover Analytics, describes an attention-guided hybrid architecture that merges MobileNetV3, DenseNet121, and the pretrained transformer model DINOv2 to distinguish the five major leukocyte categories: basophils, eosinophils, lymphocytes, monocytes, and neutrophils.
The motivation behind the work lies in a persistent difficulty in hematological image analysis. Microscopic smear images of different white blood cell subtypes can look strikingly similar, and their appearance is further altered by staining variation, differences in microscope and acquisition setup, and the presence of surrounding background content such as red blood cells, platelets, and staining artifacts. A single convolutional network tends to emphasize local texture and hierarchical spatial patterns but may miss broader contextual relationships, while a transformer alone can encode long-range interactions yet may lose the fine-grained local morphology that cell-level recognition depends on. The authors argue that neither paradigm, used alone, fully captures what is needed for reliable recognition across heterogeneous images.
Their solution is a three-branch design in which each input image is processed simultaneously by three parallel feature extractors. MobileNetV3, a lightweight convolutional network, captures efficient local texture and edge-sensitive cues. DenseNet121, with its densely connected layers, preserves and reuses hierarchical morphology features across the network’s depth. DINOv2, used strictly as a pretrained feature extractor rather than trained through its original student-teacher framework, contributes transformer-based global semantic representations through its CLS-token output. Each branch produces a 512-dimensional embedding, and these aligned vectors are then combined through feature stacking, layer normalization, and a multi-head attention block that adaptively reweights the contribution of each branch before a residual-style classification head delivers the final five-class prediction.
The model was trained and evaluated on a combined dataset built from the public PBC-Normal and Raabin-WBC collections, totaling 27,606 images after a rebalancing procedure. The original Raabin-WBC set was heavily imbalanced, with 8,891 neutrophils but only 301 basophils, so the researchers created a balanced intermediate version through augmentation and down-sampling before merging. The final combined dataset contained 3,337 basophils, 6,158 eosinophils, 4,685 lymphocytes, 4,108 monocytes, and 9,318 neutrophils. Class-weighted cross-entropy, computed only from the training split to avoid information leakage, was used to further guard against majority-class bias.
A distinctive feature of the study is its systematic treatment of image preparation. Rather than fixing a single preprocessing pipeline, the team evaluated four conditions: not cropped and minimally prepared, not cropped and processed, cropped and minimally prepared, and cropped and processed. All images were resized to 224 by 224 pixels and scaled to the [0,1] range, while the processed condition additionally applied channel-wise ImageNet normalization. Importantly, the pipeline excluded histogram equalization, CLAHE, stain masking, segmentation, denoising, and other operations that could alter stain intensity, nuclear texture, or cell boundaries, so the numerical standardization did not selectively remove any cellular or background structures.
The results showed a consistent advantage for standardized preprocessing. On the combined dataset, preprocessing lifted the proposed model’s accuracy from 94.7% to 96.7% under the not-cropped condition and from 96.3% to 97.9% under the cropped condition, with the macro-average rising correspondingly. The cropped and processed setting produced the best combined-dataset result, with class-wise F1-scores of 97.2% for basophils, 97.5% for eosinophils, 98.2% for lymphocytes, 95.8% for monocytes, and 99.3% for neutrophils. Training curves showed faster initial improvement and higher terminal accuracy under processed conditions across the compared models, suggesting that controlled numerical conditioning contributes to more stable learning.
The proposed architecture was benchmarked against four baselines trained and evaluated under the same protocol: ConcatNeXt, a fusion-oriented comparison model; the Swin Transformer; standalone DINOv2; and standalone DenseNet121. ConcatNeXt was the strongest baseline in several settings, reaching 96.2% accuracy under the cropped and processed condition, while Swin Transformer achieved the best baseline macro-F1 at 94.4%. The proposed model exceeded ConcatNeXt’s accuracy by 1.7 percentage points and Swin Transformer’s macro-F1 by 3.2 percentage points, with the gap particularly evident for monocytes, where the hybrid model reached 95.8% F1 compared with 89.4% for ConcatNeXt and 90.5% for Swin Transformer. DenseNet121 alone showed the weakest and most variable results, indicating greater sensitivity to dataset heterogeneity.
Dataset-specific analyses reinforced the pattern. On PBC-Normal, the proposed model achieved precision values ranging from 98.2% to 99.8%, recall values from 98.6% to 100.0%, and F1-scores between 98.8% and 99.4% across the five classes, yielding a macro-average of 99.1% and overall accuracy of 99.3%. DINOv2 was the strongest baseline there at 97.3% macro-F1. On Raabin-WBC, the proposed model recorded the highest macro-average F1 at 97.4%, with relatively uniform class-wise scores, ahead of ConcatNeXt at 94.9%, DINOv2 at roughly 93.4%, Swin Transformer at 92.6%, and DenseNet121 at 90.9%. Class-wise F1 heatmaps further showed that the hybrid model maintained balanced performance across all five categories rather than excelling in only a subset.
The authors are careful to frame these numbers within their limitations. The benchmark covers only five mature leukocyte categories and may not represent the full range of staining protocols, microscope configurations, and population-level variability encountered in wider laboratory practice. Robustness was assessed within the controlled image-preparation conditions rather than under broader perturbations or distribution shifts, and the study did not examine model calibration, uncertainty estimation, or systematic domain-shift sensitivity. The three-branch architecture is also inherently more complex than a single-backbone alternative, and the authors note that parameter efficiency, inference-time cost, and deployment-oriented simplification warrant further investigation.
Future directions outlined in the paper include external validation on independently collected datasets, extension to finer-grained hematological subtypes, domain generalization across different staining and acquisition conditions, and efficiency-oriented refinements such as compression and pruning, alongside uncertainty-aware prediction and interpretability studies. Within its tested scope, however, the study offers a clear empirical message: how an image is framed and numerically prepared is not an incidental implementation detail but a meaningful experimental factor, and combining complementary feature families through attention-guided fusion can outperform any single representation source alone in the automated recognition of white blood cells.
Subject of Research: Deep learning classification of five white blood cell types from microscopic smear images
Article Title: Attention guided hybrid fusion of MobileNetV3 DenseNet121 and DINOv2 for five class white blood cell classification
Article References: Shehan, M. S. H., Islam, M., Rahim, A., Uddin, M., Rahman, M. M., Rana, M. S., Islam, M. A., Bakhsh, M. M., Kabtia, M., & Raihan, D. I. (2026). Attention guided hybrid fusion of MobileNetV3 DenseNet121 and DINOv2 for five class white blood cell classification. Discover Analytics, 4(1), Article 37. https://doi.org/10.1007/s44257-026-00086-1
Image Credits: AI Generated
DOI: 10.1007/s44257-026-00086-1
Keywords: white blood cells, deep learning, DINOv2, DenseNet121, MobileNetV3, attention fusion, hematology, medical imaging, image classification, preprocessing, Raabin-WBC, PBC-Normal
News Source: Denise Maddox. (October 29, 2026). Three-Branch AI Fusion Reaches 99.3% Accuracy in White Blood Cell Classification. Scienmag.
