Deep learning and uncertainty estimation in bacterial colony analysis

This project focused on applying deep learning and statistical uncertainty estimation to intelligent models designed for counting and classifying bacterial colonies in Petri dishes. It combined methods from artificial neural networks with statistical techniques previously developed in the analysis of scattering data.

The main objective was to propose a reliable framework for estimating the uncertainty of predictions produced by deep-learning models used in biomedical image analysis. Particular attention was devoted to U-Net and U2-Net architectures, widely used for image segmentation and object detection tasks.

The main outcome of the project was the development of modifications to the U2-Net architecture that significantly improved counting precision, even in images containing a large number and a high diversity of objects. These results led to a publication in Scientific Reports and demonstrated that combining modern deep-learning approaches with robust statistical methods can substantially improve the reliability and practical value of automated image-analysis systems.

Funding: Mozart programme, Wrocław Centre for Academic Cooperation (WCA)

Industry partner: NeuroSYS, Wrocław

References:
Sci Rep 12, 10583 (2022)

Collaborators: Tomasz Golan, Jarosław Pawłowski, Sylwia Majchrowska

Figure: [place for illustration]