WDSSNet: Wavelet-decoupled selective state learning for fine-grained microbial microscopy image classification

Fine-grained microbial microscopy image classification is difficult because diagnostic contours and internal textures are subtle, while scale, focus, illumination, and background vary within a class. This study presents WDSSNet, a wavelet-decoupled selective state network that combines four Wavelet-based Purification Units (WPUs) with a Wavelet-enhanced Selective State Module (WSSM). Each WPU separates an approximation subband from three directional detail subbands, processes them independently, and applies image-level orientation weighting. The WSSM couples four-direction state-space modelling with feature reconstruction from raw wavelet coefficients. Experiments used fixed train/validation/test partitions of the Environmental Microscopy Image Collection (EMIC; 14 diatom genera; 4460 images), EMDS-6 (21 classes; 840 images), and MD-AS-2025 (five activated-sludge taxa; 14,257 images), with five independent runs per model. WDSSNet achieved 93.03 ± 0.59% accuracy and 92.72 ± 0.73% macro-F1 on EMIC, 72.22 ± 0.98% accuracy and 71.45 ± 1.60% macro-F1 on EMDS-6, and 96.92 ± 0.53% accuracy and 95.57 ± 0.59% macro-F1 on MD-AS-2025. Paired two-sided comparisons with DenseNet-121 yielded positive mean test macro-F1 differences on all three datasets, and all three Holm-adjusted p values were ≤ 0.0253. Dataset-level analysis showed dominant low-frequency energy together with dataset- and class-associated directional variation. These results support WDSSNet for the evaluated within-dataset settings, but do not establish cross-facility robustness or closed-loop process control.