Liam Chalcroft

Liam Chalcroft

Founding Computer Vision Scientist, Prospectral
PhD in Machine Learning, University College London
London, UK

I build imaging models that hold up outside the data they were trained on. Most of my research has been on brain lesion segmentation in routine clinical scans, where contrast, resolution and artefacts vary far more than in curated benchmarks. I now work on the same problem in spectral imaging.

Research

My PhD, supervised by John Ashburner and Cathy Price, asked how a stroke lesion segmentation model can be made to work on scans it has never seen: a different scanner, a different sequence, a resolution nobody would choose for research. The answer that held up was to stop training on real images. Generating training data from tissue labels under MRI physics constraints, rather than augmenting a fixed dataset, produced models that transfer to out-of-domain clinical data without sequence-specific retraining.

At Prospectral I lead machine learning and its integration into the wider product, applying the same generalisation questions to spectral imaging, where the physics is better specified and the labelled data is scarcer still.

  • Domain generalisation
  • Synthetic data
  • Self-supervised learning
  • Generative modelling
  • Physics-informed AI
  • Spectral imaging
  • 3D medical image segmentation

Selected publications

All publications Software CV