Machine learning augmented hierarchical tomographic image reconstruction of human organs

1) Brief description of the context of the research including potential impact

Imaging of organs at multiple resolutions can be extremely useful for understanding their function, especially if we can correlate cellular level information to whole organ processes. X-ray tomography is one of the methods that can be used to image at cellular level information; however, for a scan of an intact organ, normally this can only be achieved when the sample is small such as mouse organs (millimetres in size).

For human organs, traditionally only biopsies could be scanned at such resolution. These biopsies distort, making correlation of the high-resolution results to low-resolution images challenging.

A new technique, Hierarchical Phase-Contrast Tomography (HiP-CT) has been co-developed by UCL and ESRF, that can achieve cellular level resolution for volumes of interest (VOI) in an intact organ, starting from a full organ scan at lower resolution and hierarchically zoom into VOI at higher resolutions. This made it possible to have an intact organ and zoom in at any part to study its cellular level functions. (see human-organ-atlas.esrf.eu and mecheng.ucl.ac.uk/hip-ct)

The tomographic images obtained by HiP-CT provide information across a range of scales. This project will apply and develop machine learning algorithms use the information from all scan resolutions to obtain better feature extraction than can be obtained from a single resolution. If successful, this could enable super-resolution imaging, with a final goal of informing clinical radiology.

2) Aims and Objectives

The aim of the project is to apply and develop machine learning algorithms to assist the reconstruction of high-resolution VOIs using data from lower resolution scan of the entire organ. The objectives are as follows:

To investigate the use of generative model such as super resolution and generative adversarial nets to assist the reconstruction of VOI.

To investigate the use of unsupervised learning algorithms to identify feature points to be used in the reconstruction process.

To investigate the use of one model for all organ reconstructions or different models for different organs or organ groups.

Considering the possible limitation on data variation, to investigate potential use of data from other sources to help identify key features via unsupervised learning algorithms.

To compare the reconstruction results from the machine learning models with the reconstruction results obtained using the reference scan method. 3) Novelty of Research Methodology

HiP-CT has only been recently developed, the only method to remove noise during the reconstruction process for extreme off-axis scans is via a reference scan and traditional denoising algorithms. The use of machine learning techniques to reconstruct utilising data from a lower resolution scan is yet to be investigated.

4) Alignment to EPSRC's strategies and research areas

The project falls within EPSRC research area of engineering. The project aims to develop machine learning techniques to assist the reconstruction of HiP-CT scans, which helps towards the understanding of human organ structures. It aligns with the EPSRC's strategic priorities of artificial intelligence, digitisation and data, and also transforming health and healthcare.

5) Any companies or collaborators involved Dr. Paul Tafforeau from the European Synchrotron Radiation Facility (ESRF).