A laboratory team at the School of Medicine in Cardiff University is focused on researching methods to boost the immune system’s ability to recognise and kill cancer cells, with a focus on the augmentation of High Endothelial Venules (HEVs) to promote tumour control. The team uses Light Sheet Fluorescence Microscopy (LSFM) to obtain detailed 3D volume image analysis of these vessel networks within lymph nodes.

Transforming the 3D volume image into the topology map needed to derive key vessel statistical measures depends upon the staining process used to highlight endothelial cells in the vessel walls. The analysis of the staining process results in images with voxels that store 3D graphic data. The problem arises in this staining process in which the number of well-labelled vessels is highly variable, resulting in ‘holes’ in the vessel walls and very faint labelling of the finer capillaries within the network. To complete the segmentation of the incomplete vessels, these ‘holes’ and finer capillaries’ voxels must be filled. The basic idea is to use probes to count the number of lit voxels around these areas, and if this count reaches a set threshold parameter, the voxel being analysed is considered within the vessel network and is subsequently lit. Currently, the parameters; ‘the number of probes that encounter a lit voxel’ (D) and ‘if this count exceeds a threshold value’ (T) are manually entered data points performed by a member of the research team. With the number of iterations of different combinations of parameters and the number of voxels that need to be filled, this proves to be a time-consuming process. This project aims to increase the speed at which this segmentation phase is completed in the vessel network analysis. Therefore, improving the throughput of the current 3D analysis pipeline.

Machine learning techniques will be explored to automate the setting of the parameters T and D in the segmentation process. This will involve examining existing classical machine learning techniques based upon segmentation as well as examining existing deep learning models for similar problems. Then adapting the research into these solutions into a prototype specific for this project. For any deep learning methodology to be used, it would be essential to identify if there are suitable large datasets to train an image analysis model.