Development of new mathematical models and algorithms for analysis of 3D images with applications to monitoring of stents

Lead Research Organisation: University of Liverpool
Department Name: Mathematical Sciences

Abstract

This EPSRC iCASE studentship project sits in the EPSRC strategic areas of Numerical Analysis and Non-linear systems, in the themes of Healthcare technologies and Mathematical Sciences. It will study image analysis problems, arising from Healthcare applications, by developing and using new and advanced mathematical models and algorithm. It is motivated by the challenges of analysing and tackling CT images that have noise and streaking artefacts, which render current models fail to track both an object (organ) and the metal object (stent). Many new ideas will be investigated in the project, aiming to (i) remove or reduce the influence of noise and streaking artefacts so that existing models might work; (ii) identify the concerned organs using geometry and shapes information; (iii) track the organ changes by employing image registration ideas; (iv) assess the feasibility of Deep Learning (AI) for the segmentation task in the presence of noise and streaking artefacts.
Mathematically, the primary focus will be on mathematical development and analysis of accurate, variational, selective models that can take in prior information and track changes. Several imaging problems are studied and include segmentation, registration and fusion of images. Our models will aim to deal with texture and intensity inhomogeneity, as well as irregular patterns and metal artefact reduction. To tackle streaking artefacts of CTs due to metal objects, we consider two approaches: one to use artificial intelligence or geometry to identify organs and the other to remove or reduce such artefacts by re-analysis and improvement of the tomography models that lead to such artefacts in the first place. The noise adds extra levels of difficulty to the outstanding challenges of segmenting objects reliably which will be tackled by the so-called domain methods. Finally since our supervision team has clinicians from the Royal Liverpool University Hospital, we shall design tests to validate our models during the project. This will ensure that our imaging methods of automatic analyis and colligate treatment / disease progression will be useful to optimising treatment planning as well as monitoring, one specific application being the treatment of abdominal aortic aneurysms by endovascular sealing using a range of imaging modalities (CT, ultrasound, MR). The methodologies to be developed will be useful to a wider class of applications.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/R512011/1 01/10/2017 30/09/2022
1945983 Studentship EP/R512011/1 01/10/2017 30/09/2021 Liam Burrows
 
Description The main focus of this work is to develop mathematical models for image analysis, particularly focusing on the medical application of assisting with endovascular treatment of abdominal aortic aneurysms (AAA). The main focus has been on image segmentation, which is the aim of identifying particular objects in an image. To develop an algorithm to automatically interpret CT scans of patients with AAAs, segmenting the abdominal aorta is the most challenging part due to streaking artifacts and low contrast at the boundary.

So far, we have completed two major pieces of work:
The first is a two-stage method which allows us to segment two objects with a single user input. Selective segmentation usually require a user to input some points on an image to indicate the object they wish to segment. With this piece of work, we are able to segment two objects with a single click. This was developed for the particular application of segmenting the abdominal aorta and stent (which is usually inserted to the aneurysm as part of endovascular treatment), in order to make it easy to segment both the stent and abdominal aorta with less user interaction. Although we initially developed it for the stent and aorta problem, this piece of work has potential applications in many medical settings, such as segmenting a tumour and the affected organ, for example.
The second piece of work is an image enhancement method which aims to enhance edges in an image. A major challenge of segmenting the abdominal aorta is that it is difficult to detect an edge at the boundary, due to nearby touching objects being a very similar intensity. We aim to decompose an image to the smooth parts and edge parts, and by modelling the edges by a collection of approximated Heaviside functions, we are able to obtain a more detailed edge set, able to detect incredibly weak edges, out performing the typically used image gradient. This piece of work was a collaboration with Professor Weihong Guo from Case Western University, Cleveland, USA, who has expertise in such image enhancement methods.

We are currently in the process of developing deep learning methods to segment the abdominal aorta and stent. We are working with experienced clinicians at the Liverpool Royal Hospital who are manually segmenting datasets for us in order for us train networks to do so automatically. These neural networks should be able to segment the abdominal despite the presence of streaking metal artifacts and low contrast at the boundary of the abdominal aorta.
Exploitation Route We are developing tools to aid Royal Liverpool Hospital with their treatment of abdominal aortic aneurysms via endovascular sealing. Being able to automatically interpret CT scans will reduce the workload of clinicians, who can instead spend their time treating patients instead of in front of a computer screen.
The annotated datasets we are acquiring can be put to use in future by LCMH members to train networks or for testing their own algorithms.
Sectors Healthcare

 
Description Royal Liverpool Vascular Department 
Organisation Royal Liverpool University Hospital
Country United Kingdom 
Sector Hospitals 
PI Contribution We have developed mathematical models to assist with the analysis of CT scans.
Collaborator Contribution They have provided assistance and described the technical challenges that medical image analysis presents. In addition, they provide ground truth annotations for us to train networks, and for us to assess the performance of our models.
Impact One conference paper: https://doi.org/10.1007/978-3-030-39343-4_17 Two papers submitted for review as of March 2020.
Start Year 2017