Morpho-spectral YOLOv10-HSNet: lightweight hybrid model for multi-class leukocyte detection via attention-guided signal analysis

The analysis of microscopic images of white blood cells (WBCs) poses a complex signal processing challenge due to the presence of subtle morphological cues and significant spectral variability arising from staining protocols. This paper presents YOLOv10-HSNet, a lightweight and hybrid neural architecture for real-time, multi-class detection and classification of 12 morphologically distinct leukocyte types. The model incorporates two synergistic modules: MHSA-Morph, a multi-head self-attention mechanism guided by structural morphology to enhance spatial pattern discrimination, and Spectro-Clinical Fusion (SCF), which performs spectral signal fusion across RGB, CIELab, and HSV color domains to increase robustness against chromatic distortions and optical variability. To reinforce generalization in diverse visual conditions, the training pipeline includes simulated spectral perturbations and stochastic augmentation of signal textures. The model demonstrates improved accuracy in morphological signal decoding, particularly for rare or noisy cell categories such as smudge cells, myelocytes, and blasts. Experimental results show that YOLOv10-HSNet surpasses the baseline YOLOv10, improving accuracy from 85.00 to 88.21% (+ 3.21 pp), F1-score from 84.00% to 87.55% (+ 3.55 pp), precision from 82.58 to 91.31% (+ 8.73 pp), and recall from 79.64 to 84.65% (+ 5.01 pp). These gains validate the model’s effectiveness in learning fine-grained morphological signals and handling spectral-domain inconsistencies.