1: \begin{abstract}

2: Diffusion models are a special type of generative model, capable of synthesising new data from a learnt distribution. 

3: We introduce \modelname, a diffusion-based model for solving the inverse problem of three-dimensional (3D) cell shape prediction from two-dimensional (2D) single cell microscopy images. 

4: Using the 2D microscopy image as a prior, \modelname\ is conditioned to predict realistic  3D shape reconstructions.

5: %Since diffusion models rely on a stochastic sampling process that allows to generate infinitely many predictions, we exploit this process to generate five reconstructions of every microscopy image.

6: To showcase the applicability of \modelname\ as a data augmentation tool in a feature-based single cell classification task, we extract morphological features from the red blood cells grouped into six highly imbalanced classes. 

7: Adding features from the \modelname\ predictions to the three minority classes improved the macro F1 score from $F1_\text{macro} = 55.2 \pm 4.6\%$ to $F1_\text{macro} = 72.2 \pm 4.9\%$.  

8: We thus demonstrate that diffusion models can be successfully applied to inverse biomedical problems, and that they learn to reconstruct 3D shapes with realistic morphological features from 2D microscopy images.

9: \end{abstract}

10: