Spatial Atlas of Lung Tissue Fronts: Deep Learning–Enabled Segmentation and Microenvironmental Gradient Analysis
Code: BBSRC-DFA_2026_23 (CASE)
Primary Supervisor: Prof Greg Slabaugh
Email: g.slabaugh@qmul.ac.uk
Institute: Digital Environment Research Institute
Secondary Supervisor: Dianne Cooper
Email: d.cooper@qmul.ac.uk
Institute: William Harvey Research Institute
CASE Partner: AstraZeneca
Abstract:
High-plex spatial transcriptomic platforms generate single-cell-resolved molecular profiles within intact tissue architecture, producing multimodal datasets that combine transcript coordinates, morphology, and expert annotations. However, computational approaches for robust segmentation of structurally heterogeneous tissue and for modelling continuous spatial transitions across complex interfaces remain limited. Conventional deep learning pipelines based on convolutional neural networks struggle to capture long-range spatial dependencies and irregular spatial relationships inherent to image-based spatial biology data.
This project will develop advanced deep learning methodologies for multimodal segmentation and spatial gradient modelling in high- high-resolution spatial lung tissue datasets exhibiting structural heterogeneity . Using spatial transcriptomics integrated with matching H&E histology and pathologist-defined regional annotations, the student will design architectures that combine state-space and attention-based mechanisms to model long-range spatial dependencies across heterogeneous tissue regions. The work will emphasise algorithmic innovation in multimodal fusion, integrating transcript-level coordinates, cell-state information, and tissue/cell morphological features within unified spatial representations.
Building on these representations, spatial graph neural networks will be developed to characterise cell-cell and cell-matrix neighbourhood structure, enabling quantitative modelling of continuous microenvironmental gradients at tissue interfaces. The project will further explore continuous spatial progression modelling frameworks to infer structured transitions across heterogeneous regions from cross-sectional data.
The resulting methods will be implemented as reproducible, open-source software for multimodal spatial image analysis and will generalise to other high-plex spatial imaging contexts involving complex tissue interfaces. By advancing machine learning approaches for segmentation, graph-based modelling, and gradient inference in spatial biology, the project directly aligns with the objectives of the AI for Multi-modal Spatial Doctoral Focal Award.
Lay Summary:
New imaging technologies can now measure the activity of thousands of genes directly inside intact tissue, while preserving its natural structure. This means researchers can see not only which cells are present, but also exactly where they are located, how they are organised and how this links to their function/behaviour. These datasets are extremely rich but also very complex, and current computer tools struggle to analyse them accurately.
Tissues are not uniform. In complex tissues, different regions can look and function very differently from one another. Important biological changes often occur at the boundaries between tissue regions. Identifying these boundaries and understanding how individual cells behave or how groups of cells interact and change across tissue boundaries is difficult using existing software.
This PhD project will develop new artificial intelligence (AI) methods to analyse these complex spatial datasets. The student will create advanced computer models that combine gene expression measurements with standard microscope images of tissue. These models will automatically identify different tissue regions and measure how molecular signals change across space.
Although the work will use lung tissue data as an example, the tools developed will be designed to work across many types of spatial imaging studies. The outcome will be new open-source software that helps researchers better understand how cells are arranged and how they interact within intact tissues.
Aims and Objectives:
Aim 1: Develop multimodal deep learning architectures for robust segmentation of heterogeneous spatial tissue
Design state-space and attention-based segmentation models capable of capturing long-range spatial dependencies, building on prior work in the Slabaugh lab (e.g., CAMS [1]). Integrate transcript-level spatial coordinates, cell type and cell-state information, morphology, and expert annotations within unified multimodal representations. Benchmark performance against convolutional and clustering-based baselines using quantitative segmentation and generalisation metrics.
Aim 2: Construct spatial graph neural network frameworks for modelling neighbourhood structure
Represent tissue as a cell-level graph with edges defined by spatial proximity and molecular similarity. Develop Graph Neural Network architectures to learn embeddings that capture higher-order neighbourhood organisation and stable microenvironmental states across samples.
Aim 3: Develop continuous spatial gradient and progression modelling methods
Move beyond discrete region labels by constructing continuous embedding spaces that model structured spatial transitions. Explore neural differential equation or diffusion-inspired approaches to quantify gradient strength, directionality, and transition dynamics across heterogeneous interfaces.
Aim 4: Deliver generalisable, reproducible software for multimodal spatial analysis
Implement modular, version-controlled (GitHub) pipelines and release open-source software adhering to FAIR principles, enabling application beyond the exemplar lung dataset.