1st Supervisor: Dr Jack Lee, King’s College London
2nd Supervisor: Dr Andrew King, King’s College London, Imperial College Clinical Supervisor: TBC
Aims of the Project:
- To develop an AI method to reconstruct 3d coronary vessel structure from 2D angiograms
- To perform validation on a large patient dataset of angiograms and matched CTA
- To demonstrate the technology real time in the catheterisation laboratory in interventional context
Lay Summary:
Coronary artery disease is a leading cause of death in the UK. Each year, more than 250,000 diagnostic coronary x-ray angiographies are performed, costing around £1500 per procedure. However, very little of the image acquired is used for diagnosis – the typical diagnosis will involve a visual assessment of the diseased vessel which can be subjective and known to be <70% accurate. To extract more from the data, a functional analysis involving pressure is needed. In the computed tomography angiography (CTA) clinical practice, established computational methods now allow the invasive pressure assessment to be avoided, cutting down procedure time, saving on cost and reducing patient discomfort. These innovations achieved in the CTA field are not currently possible in the cathlabs, chiefly due to the inability to process the image data in an automated and efficient way.
The key characteristic that sets the x-ray angiography apart from other relevant imaging modalities is its 2D nature. During the cathlab procedure, typically only a few images are taken from different angles thus yielding a challenging sparse dataset from which to reconstruct a 3D object. The previous attempts to address this task used physics and computer vision-inspired approaches. While they are capable of producing successful results, their shortcomings include susceptibility to subject motion and image noise, and crucially, requirement of manual processing and long computation time to achieve the goal which make them unsuitable for real time clinical applications such as during a percutaneous coronary intervention.
The approach we propose is intrinsically different in that an artificial intelligence technique serves as the basis for the 2D-to-3D reconstruction. This means overcoming the technical challenges are front-loaded within the design and training stage, thereby sidestepping much of the shortcomings of the traditional approaches to allow a fast, full automation. Currently there is no published technique that achieves this goal. In our prior work we have developed an AI approach which is capable of reconstructing vascular networks accurately that was trained largely on synthetically generated 2D images. Moreover, we have set up an extensive supporting framework to manipulate 2D and 3D training data and to allow a comprehensive comparison of the potential AI approaches.
In the present project, we aim to greatly extend the functionality of our AI 3D reconstruction approach and demonstrate its clinical usefulness by performing real world data test in the cathlab setting. The outcome will allow a novel software tool that will integrate readily into the fast-paced cathlab work flow as well as further flow and pressure-estimation models and offer improved diagnosis at reduced cost and discomfort to the patients. The ideal candidate for this project will have a background in AI and machine learning, mathematics and medical image analysis, with an interest in engaging with clinicians for joint real world problem-solving.