Accurate bacterial identification is essential for effective treatment selection in human and veterinary medicine, as well as for detecting contaminants in food, water, and cosmetics. Traditional methods, such as culture-based techniques, are slow and labor-intensive, often requiring up to seven days for results and highly skilled personnel. Modern molecular approaches, like PCR, provide faster results but are costly and limited in identifying multiple bacterial species in a single test. Additionally, both approaches can be sensitive to sample contamination and require specialized reagents. Our client, a Polish startup, sought a more efficient, scalable, and accessible solution.

The project focused on developing an innovative optical system that records Fresnel patterns of bacterial colonies grown on Petri dishes. These images were processed using a custom-built algorithm to extract morphological and textural features that differentiate bacterial species. A machine learning model was designed to classify these patterns, recognizing one of dozens of species from an existing database. If the similarity to known bacteria was too low, the system generated a no-match response, allowing further investigation. This approach enabled fully automated, non-invasive bacterial identification with minimal manual intervention.

The developed system offers a rapid, cost-effective, and highly accurate bacterial identification method, seamlessly integrating into microbiology laboratories. It achieved over 96% accuracy, as confirmed by a British certified laboratory under ISO standards. The analysis time (including colonia growing) was reduced to under 24 hours, aligning with standard lab workflows. Unlike traditional methods, the system can identify multiple bacterial species in a single test while using standardized equipment and reagents. Additionally, it allows reanalysis of the same sample with alternative methods, which is not possible with conventional approaches. This solution provides a simple, automatic, and scalable tool for microbiological diagnostics across various industries.

The project covered the full product development cycle, transforming an early-stage university prototype into a compact, market-ready device. We collaborated closely with the client to address challenges common in experimental lab-based innovations, including iterative development of the bacterial identification pipeline, transitioning from catalogue strains to real-world patient and production line samples, and refining species differentiation by more and more effective algorithms. The solution evolved to recognize not only distinct bacterial species but also closely related strains, scaling up to handle hundreds of different species.
Additionally, we overcame laboratory constraints by ensuring that the incubation process aligned with microbiology lab operations, transitioning from lab-grade to mass-production components, and maintaining experimental consistency despite bacterial sample variability. The device itself underwent several iterations, improving in accuracy, size, and cost efficiency to become a commercially viable product. Each iteration required method validation, data collection, and further refinement of the identification method.
By combining cutting-edge optical technology with machine learning, this project redefined bacterial identification, offering a faster, more affordable, and non-invasive alternative to traditional microbiological methods. Now positioned as a pioneering tool for laboratories, hospitals, and food safety professionals, the system sets a new standard in bacterial detection, making high-precision diagnostics accessible on a larger scale.

The machine learning pipeline was designed to be flexible and modular, allowing the exchange of processing modules and adapting their complexity based on the device’s development phase. Image preprocessing techniques were used to enhance bacterial colony patterns, followed by feature extraction methods incorporating mathematical morphology and algorithmic approaches. Predictive machine learning models, including neural networks as benchmarks, were developed and refined to improve classification accuracy. The software was designed for process automation, balancing precision with computational efficiency to ensure seamless integration into microbiology laboratory workflows.