Development and Application of a Generative 3D Cell Morphable Model for Blood-Cell Morphodynamics

About the project

About the Project This iCASE project addresses a critical bottleneck in blood-cell morphodynamics: high-throughput blood-smear platforms capture large-scale clinical 2D cell images, but the smearing process physically flattens cells onto glass, destroying the 3D nuclear and cellular geometry that encodes mechanics and cell state. The core 3D Cell Morphable Model (3D-CMM) will be developed and deployed. Dr Shanxin Yuan will lead the generative 3D-CMM framework, along with DERI director Prof Greg Slabaugh. Co-supervisor Dr Oscar Maiques will provide biological validation data, including 3D confocal Lamin B1-labelled B-cell imaging and DLBCL tissue biomarker data, developed in partnership with Prof Andrejs Braun's BCI Genome Regulation group. The aim is to build, validate, and deploy the 3D-CMM: Objective 1: Establish the 3D shape prior and smearing deformation model from a training partition of Maiques lab Lamin B1 confocal z-stacks. A shape space learned from primary B-cell nuclei captures biologically plausible geometries. The physical smearing transformation, compressing 3D nuclear geometry into a 2D projection, is modelled as a differentiable forward model, enabling the generation of synthetic paired 2D–3D training data through virtual smearing of confocal shapes, providing supervised signal before any public data is used. Objective 2: Train the 3D-CMM on 75,000+ harmonised public images (MLL23, AML-Cytomorphology, and Acevedo peripheral blood datasets) using the confocal-derived prior. A differentiable renderer implements the smearing forward model; an analysis-by-synthesis loop optimises a disentangled latent code, encoding cell identity, nuclear shape, and deformation state, until the rendered 2D projection matches the observed image. Population consistency across thousands of cells substitutes for angular diversity unavailable in blood-smear imaging. For non-B-cell types, cell-type-specific deformation fields extend the B-cell prior using population 2D statistics, shape-from-shading, and symmetry constraints. Objective 3: Apply the validated framework to circulating B-cell and lymphoma nuclear morphology using a two-arm design. Arm 1 will estimate the 3D-CMM parameters for deformation patterns, such as lobularity, groove geometry, and chromatin texture, in circulating B cells, CLL, MCL, and leukaemic-phase lymphoma. During the industrial placement, the student evaluates the framework on large-scale real-world blood-smear data through Aimagine Care's clinical deployment network, assessing robustness, scalability, and translation to routine practice. Arm 2 benchmarks the 3D-CMM against held-out confocal Lamin B1 z-stack nuclear geometries from primary B cells and correlates scores with DLBCL Lamin B1 tissue data via the Maiques-Braun BCI collaboration, establishing a non-invasive morphological surrogate for tumour nuclear-architecture state. This fits the BBSRC remit in Rules of Life, Predictive Biology, and Technology Development. It asks how physiological activation and nuclear-lamina state shape lymphocyte nuclear geometry, and whether that geometry can act as a cell-state barcode without transcriptomic inference. Recent Braun-linked work in HemaSphere shows that Lamin B1 safeguards B-cell genome stability and shapes DLBCL outcome. By linking scalable blood-film AI to 3D confocal Lamin B1 ground truth, the project will create an open-source tool for predicting biophysical state from routine morphology.

Research areas