Advanced Computational Imaging and Generative AI for High-Fidelity modelling of Multi-Modal Spatial Biology
Code: BBSRC-DFA_2026_17
Primary Supervisor: Pengfei Fan
Email: pengfei.fan@qmul.ac.uk
Institute: Digital Environment Research Institute
Secondary Supervisor: James Timmons
Email: j.timmons@qmul.ac.uk
Institute: William Harvey Research Institute
Abstract:
Imaging-based spatial transcriptomics (e.g., Merscope) offers unprecedented potential to map subcellular gene expression within intact human tissues. However, in structurally complex tissues with a dramatic range of cell size e.g. human skeletal muscle, or layered cell structures (CNS) analysis is bottlenecked by quality assuring raw image quality and adjusting for optical noise. Heuristic cell segmentation methods can fail to resolve irregular interstitial cell shapes, leading to transcript contamination. This project aims to develop a suite of novel Generative Artificial Intelligence (AI) frameworks focused explicitly on signal restoration, morphological cell shape inference, and rigorous unsupervised benchmarking. To achieve this, we will focus on three methodological innovations. First, we will develop Physics-Informed Diffusion Models to mathematically reverse optical degradation and denoise raw Merscope Z-stack signals. Second, to improve segmentation of cells lacking optimal membrane staining, we will utilize Generative Domain Adaptation as well as multiple image stains that differentially represent the membranes from distinct cell types. The project has a unique database of single-cell resolution Merscope images, large-scale H&E images and extremely high-resolution Electron Microscopy (EM) images of the tissue vascular system. By leverage these data, the AI will learn and transfer structural shape priors to guide fluorescence image segmentation. Third, to benchmark to a ground-truth we will evolve robust frameworks utilizing the Mutually Exclusive Co-expression Rate (MECR), taking advantage of known cell-specific marker genes to rigorously validate image and segmentation improvements. This project will deliver broadly applicable, computer-science-led AI methodologies for image-based data restoration, enabling researchers to confidently extract true biological signals from noise.
Lay Summary:
To study complex physiology, such as aging, we need to see exactly where specific genes are active inside the tissue and in fact which subcellular compartment. New spatial transcriptomic technologies allow us to do this, but they generate noisy 3D images. Cells are tightly packed together such that standard methods (e.g. Cellpose) struggle to draw accurate boundaries around individual cells. This results in blending gene expression signals from neighboring cells and confusing the results. This has necessitated the “human in the loop” approach where manual correction of initial computer driven segmentation, by human hand, is required to get acceptable segmentation performance. This PhD project will address these problems by developing advanced Artificial Intelligence (AI) to act as a virtual lens that cleans and clarifies these images. The student will build AI models using first principles from physics and math to remove noise from the raw microscope data. The project will also use very high-resolution Electron Microscopy (EM) pictures of the small irregular vascular cells to improve the modelling. The AI will learn the true "shapes" of these cells and use this knowledge to automatically find them in our spatial transcriptomic Merscope images. Finally, we check if our methods are working by using strict biological rules—checking that genes which should never naturally exist in the same cell, are indeed properly separated by our models.
Aims and Objectives:
- Aim 1 (Restoration & Fusion): Develop Physics-Informed Diffusion Models and self-supervised architectures to automatically select, fuse, and denoise 3D Z-stacks, while correcting physical sectioning artefacts via generative inpainting.
- Aim 2 (Domain Adaptation): Develop cross-modal Generative Adversarial Networks (GANs) to transfer ultrastructural "shape priors" from a database of 1,800 high-resolution Electron Microscopy (EM) images to low-resolution MERSCOPE fluorescence data, enabling accurate segmentation of interstitial cells.
- Aim 3 (Unsupervised Benchmarking): Establish a robust, ground-truth-free validation framework utilizing the Mutually Exclusive Co-expression Rate (MECR) to mathematically benchmark AI improvements in image quality and transcript assignment.