Water is vulnerable to waterborne diseases caused by pathogenic microorganisms posing health risks to both humans and animals. The efficient detection of these contaminants in large water volumes remains a challenge due to their presence in small concentrations and also mixing with other particles. The development of a rapid and highly sensitive detection system operating in-line with the water chain is essential for enhancing water safety and public health protection. To address this, we present a novel biosensing platform combining miniaturized Total Internal Reflection Fluorescence (TIRF) microscopy with machine learning (ML) for rapid, sensitive detection of E. coli and MS2 coliphages in wastewater. Preliminary tests with sub-micron fluorescent particles confirmed a high signal-to-noise ratio. The system integrates a CMOS camera for fluorescence and bright-field imaging with a photodiode for simultaneous emission intensity measurements, generating complementary datasets for real-time analysis.

The sensing approach uses antibody-functionalized surfaces to selectively capture target microorganisms. Prior to photoluminescence mapping, a control platform of UV-absorbance measurement on a plate reader was adopted to evaluate the reproducibility of functionalization on identical substrate surfaces. The findings further confirmed the selectivity and reproducibility of the functionalization process on this platform. Silicon, BK7, and UV-fused silica substrates were investigated, with micro-photoluminescence mapping showing that BK7 and fused silica provide uniform and reproducible functionalization. In combination with fluorescently labelled secondary antibodies, these surfaces support highly specific sandwich immunoassays, enabling reliable visualization and quantification of bound pathogens at the TIRF interface. This configuration ensures high sensitivity while minimizing background noise from non-specific particles.

A tailored ML pipeline was developed to process microscopy images for pathogen detection and quantification. Using curated datasets from open-source repositories and laboratory experiments, several models were evaluated, with YOLOv12 achieving the best balance of precision, recall, and mAP@0.5. Interpretability was enhanced using Grad-CAM, which highlighted image regions driving predictions. Together, these results demonstrate a compact, explainable, and highly sensitive biosensing system with strong potential for automated, real-time water quality monitoring and improved public health protection.