Enhanced Publication 260124 | Lightweight Spatial-Spectral Dual Branch Transformer Network for Classifying Polarized White Blood Cell Hyperspectral Images

However, existing methods based on stained images are susceptible to variations in staining conditions and imaging techniques, while traditional hyperspectral imaging remains limited in distinguishing white blood cell subtypes with similar morphological and spectral characteristics. To address this, this study proposes a research approach combining polarized hyperspectral imaging (PHSI) technology with deep learning. It validates the reliability of polarized hyperspectral imaging in white blood cell classification tasks, enabling precise identification of white blood cells and providing an efficient and reliable technical solution for clinical diagnostic support.  Methods  This study first established a polarized hyperspectral microscopic imaging system.

By controlling the polarizer at different angles, multiple polarized images were acquired. Based on Stokes vector theory, polarization parameters were calculated to obtain a multidimensional data cube containing both light intensity and polarization state information. Furthermore, linear polarization degree images were computed, thereby constructing a polarized hyperspectral image dataset of white blood cells.

Given the multidimensional characteristics of PHSI data, a lightweight spatial−spectral dual−branch Transformer network (LSDBT) is proposed. Preprocessed data are first fed into a dual−branch feature extraction module, which separately extracts local spatial features and spatial−spectral features. An adaptive scaling factor is introduced to fuse these two feature types, thereby fully capturing the multidimensional information inherent in PHSI.

The network backbone employs a lightweight Swin Transformer, streamlining the feature learning process to achieve effective global modelling while reducing computational complexity. The model terminates with a classification output via global pooling and a fully connected layer. Its performance in white blood cell classification is systematically evaluated using core metrics: overall accuracy, precision, recall, and F1 score.

The network's efficacy in classifying polarized hyperspectral white blood cell images is validated through designed ablation studies, comparative experiments, and feature visualization.  Results and Discussions  The visualization results of DOLP spectral lines and PHSI for white blood cells reflect the selective absorption and scattering of light by the internal components of monocytes, lymphocytes, and neutrophils.

It is evident that the polarization characteristics differ significantly across cell types, and polarized hyperspectral imaging effectively enhances the image contrast of white blood cells, providing polarization information in the spatial dimension that distinguishes it from traditional intensity data. Through experiments with LSDBT, the network achieved an overall accuracy of 99.29% on the test set, with balanced and relatively superior evaluation metrics across all categories.

This study experimentally investigated the scaling factor within the module, which adjusts the output feature intensity of the spatial−spectral branch. Results indicate that model performance initially increases then stabilizes with rising scaling factors, suggesting that appropriately increasing the weight of spatial−spectral features enhances classification performance. The scaling factor was ultimately set to 3.

Ablation experiments confirmed the significant contribution of the dual−branch feature extraction module to performance enhancement. Although the lightweight design incurred minor losses, it substantially reduced computational complexity and model parameter count. Compared to hyperspectral imaging, PHSI data demonstrated higher accuracy across all classifiers, proving that polarization information provides unique physical characteristics that enhance cell distinguishability.

When benchmarked against existing state−of−the−art models, LSDBT achieved optimal performance across multiple metrics. Furthermore, t−SNE results reveal that the features extracted by this network exhibit high clustering and distinct inter−class separation. The computational costs of different models were compared, and although LSDBT does not have the lowest computational cost, it achieves the best classification performance while maintaining a balance between model size and computational efficiency.  Conclusions  This study proposes a lightweight dual−branch Transformer network for polarized hyperspectral white blood cell classification, pioneering the integration of polarized hyperspectral imaging with deep learning for white blood cell categorization.

Comparative analysis with existing hyperspectral datasets validates the superior performance of this technique in white blood cell classification tasks. In terms of classification model design, the proposed LSDBT network integrates spatial and spectral information through a dual−branch feature extraction module, achieving efficient modelling and classification via lightweight operations. Results demonstrate that this design achieves high feature discriminative power and distinctiveness while reducing computational complexity, yielding excellent performance in polarized hyperspectral white blood cell image classification.

This not only provides a high−precision, lightweight method for automated white blood cell classification but also offers further support for the application of polarized hyperspectral imaging technology in cellular microscopic analysis and clinical auxiliary diagnosis.