MambaPath-Net: Selective State Space Model for Efficient Waterborne Pathogen Detection in Microscopy Images

Pathogen detection from image microscopy in water is still challenging because of the morphological similarity between pathogens and the sample to sample variations. Current CNN-based approaches fail to capture long-range spatial context, and vision transformers demand high computational complexity and hence are hard to deploy in resource-restricted environments. In this paper, we introduce MambaPath-Net, a selective state space model that can capture global contextual features at linearly-computational cost. The model was designed to include multi-scale state aggregation and a pathogen-specific scanning procedure to differentiate fine-grained morphological features from eight classes of waterborne pathogens. The results demonstrate that MambaPath-Net reaches 97.34% accuracy and 96.69% F1-score, surpassing recent CNN, transformer and SSM baselines with just 14.7M parameters and 2.41 GFLOPs.