Deep Learning Applied to Plain Abdominal Radiographic Surveillance After Endovascular Aneurysm Repair (EVAR) of Abdominal Aortic Aneurysm (AAA)

Overview

Deep learning applied to plain abdominal radiographic surveillance after Endovascular Aneurysm Repair (EVAR) of Abdominal Aortic Aneurysm (AAA).

Abdominal aortic aneurysm (AAA) is a condition in which the abdominal aorta, a large artery, dilates gradually, secondary to a degenerative process within its wall. This can lead to rupture of the weakened wall with subsequent exsanguination into the abdomen. This scenario is usually fatal. The diameter of the aneurysm positively correlates with the risk of rupture. Aneurysm size is therefore the primary determinant when considering whether or not to electively repair AAAs. Endovascular aneurysm repair (EVAR) has become the standard treatment for AAAs in the vast majority of patients. It is a minimally invasive technique that aims to exclude the aneurysm from the circulation by placement of a synthetic "stent-graft" within the aortic lumen. Metallic barbs as well as radial force maintain stent-graft position in non-aneurysmal aorta above the aneurysm as well as in the iliac arteries below the aneurysm. Level 1 evidence has consistently demonstrated improved perioperative survival with EVAR as compared to traditional open surgery. However, there are concerns regarding the long-term durability of EVAR stent-grafts, with 1 in 5 patients requiring further surgery to the aneurysm in the 5 years after the operation. This is often due to failure of the position and integrity of the stent-graft. Therefore, standard international practice is to keep patients are life-long surveillance after EVAR. This is usually in the form of plain radiographs in combination with either computerised tomography (CT) or duplex ultrasound scans, all performed on an annual basis. Stent-grafts are visible on plain radiographs of the abdomen and by comparing series of images taken over time, it is possible to diagnose a myriad of stent-graft problems including migration, disintegration and distortion. But these changes can be subtle on plain radiographs and difficult to spot, even to the most trained human eye. As a result, patients undergo more detailed scans that unfortunately carry a risk of nephrotoxicity and radiation-induced malignancy. The aim of our research is to improve the diagnostic potential of plain radiographs by applying modern deep learning computer algorithms for interpretation. Artificial intelligence (AI) in the form of deep learning has shown great success in recent years on numerous challenging problems. The success of deep learning is largely underpinned by advances in powerful graphics processing units (GPUs). GPUs enable us to speed up training algorithms by orders of magnitude, bringing run-times of weeks down to days. Our study will explore the use of artificial intelligence in interpreting series of anonymised plain radiographs to identify features of a failing stent-graft. A deep-learning algorithm will be applied to post-EVAR plain radiographs that have already been performed at our institution in England over the last 10 years. We will then compare the effectiveness of the machine in identifying stent-graft related problems to the known outcomes identified by human interpretation previously. This project will rely on recent advances in deep learning techniques. It is expected that deep learning will bring good performance for EVAR surveillance in line with its successful application in domains such as the recognition of digits, Chinese characters, and traffic signs where computers have produced better accuracy than humans.

Conditions

Eligibility

Locations (1)

Outcomes