Optimising pipeline configuration through an interactive machine learning approach.

The conventional approach to optimising a set of cell profiling parameters (or ‘configuration’) requires the user to change multiple settings in a trial and error manner. This is a slow and tedious process, with quality of the image processing pipelines usually only measured after analysis of the entire dataset. Our proposed approach combines machine learning with explicit definition of quality of a pipeline configuration (or the quality score (QS)) obtained in real time. The burden of choosing a pipeline configuration is then placed on a machine learning algorithm called Bayesian optimisation (BO), which learns the optimum pipeline settings that maximises the QS. Through this interactive machine learning approach, cell profiling can be rapidly optimised, reduce cognitive load on users and ensure high quality outcomes.