Deep Learning

Together with Dr Ben Sherlock and Dr Fabrice Gielen1, I developed a DNN capable of real-time sorting for microfluidic setups with appropriate architectures. The experimental setup used microfluidic channels for the sorting of micro-droplets, such that specific classes of droplets could be selected when the experiment is run, despite a probabilistic initial droplet preparation mechanism.

Even though we lacked prior traninig data, we were able to select an AI-based approach here as the experimental setup allowed for the generation of specific classes of droplets, specifically the empty class. This meant we were faced with an imbalanced model-training problem.

To tackle this, we used a combination of convolutional layers and max-pooling, resulting in a lean network architecture, while still achieving excellent accuracies for distinguishing empty from not-empty droplets.

In addition, manual annotation for the individual classes together with data augmentation, allowed us to extend the capabilities to the distinction between the various classes of non_empty droplets, once the initial network had pre-sorted the data.



As part of work with Dr Rikke Morrish, I developed a image analysis pipeline capable of detecting and tracking cells in Microfluidic channels and quantizing their deformation2. The development relied on lower-level functionality provided by the Scikit Image library, together with a single-particle tracking algorithm.

While this project relied on image-analysis experience and expertise instead of the large quantities of training data typically required for ML-based approaches, we achieved very favourable analysis results compared with the laborious manual segmentation that had been used previously. This demonstrated that for certain applications, particularly novel experiments which may not be well suited to the ML/AI paradigm due to entirely new data, such classical image processing approaches can be extremely valuable for both generating research quality results, and additionally as a mechanism to generate training data in a semi-automated way.