Liam Chalcroft
Liam Chalcroft

PhD, University College London
London · Email

I'm a machine learning researcher working on medical and spectral imaging. I'm currently the founding computer vision scientist at Prospectral, where I lead the machine learning work.

My PhD at University College London studied stroke lesion segmentation in clinical MRI, focusing on robustness across scanners and sequences.

Tokenizer-Generator Coupling in Medical Image Generation

NeurIPS 2026, poster · Liam Chalcroft

The study compares tokenizers, generators and sampler settings on ChestMNIST images at 64 by 64 resolution.

In this study, tokenizer rankings depend on the generator and sampler settings. Reconstruction PSNR alone is not a reliable criterion for selecting a tokenizer.

Unconditional 64 × 64 images, evaluated with non-clinical generation metrics.

Original two-panel tokenizer-generator comparison: seed variation and FID measurements across generator families.
Tokenizer-generator comparison. Results for a vocabulary of 1,024 tokens compare three training seeds; the remaining configurations use one seed. Source and study scope · Full-size figure

GAZE

Duaa Alim, Mogtaba Alim and Liam Chalcroft
AIiH 2026, oral paper

GAZE lets a vision-language model inspect brain MRI iteratively, using image tools and literature retrieval. It evaluates diagnosis, localisation and captioning together on cases covering rare neurological conditions.

Retrieval effects depend on the model: gains in diagnosis can coincide with losses in localisation. The paper examines these trade-offs alongside structured prompting and tool use.

Tools and evaluation in GAZE

  • Viewer tools

    The model can request zoom, windowing, contrast adjustment and edge detection.

  • Retrieval tools

    The model can search PubMed for literature and Open-i for radiological images.

  • Joint evaluation

    Structured outputs are scored jointly for diagnosis, localisation and captioning.

Method overview from the public workflow documentation and paper.

This paper studies how to schedule physics guidance in diffusion models for photonic inverse design.

Gradient-manifold alignment scheduling for physics-guided diffusion

SPIE Photonics Europe 2026

  1. Domain-Agnostic Stroke Lesion Segmentation Using Physics-Constrained Synthetic Data

    L. Chalcroft, J. Crinion, C.J. Price, J. Ashburner.

    MICCAI 2025.

    2025

  2. Synthetic Data for Robust Stroke Segmentation

    L. Chalcroft, I. Pappas, C.J. Price, J. Ashburner.

    MELBA 2025.

    2025

  3. DeepISLES: a clinically validated ischemic stroke segmentation model from the ISLES'22 challenge

    E. de la Rosa, M. Reyes, S.-L. Liew, et al. Including L. Chalcroft.

    Nature Comms 2025.

    2025