A software named PRAEVAorta (Nurea), using artificial intelligence (AI), has the potential to enable a fast, reproducible, and fully automated analysis of abdominal aortic aneurysm (AAA) sac pre- and post-endovascular aneurysm repair (EVAR). This is according to Caroline Caradu (Bordeaux University Hospital, Bordeaux, France) and colleagues, who recently reported the results of a single-centre, retrospective analysis of fully automatic volume segmentation using deep learning approaches in the Journal of Vascular Surgery (JVS).
Precise follow-up remains the “Achilles’ heel” of EVAR, the researchers note in the introduction to their report, detailing that the procedure’s long-term success has been “hindered” by postoperative complications such as endoleaks, endotension, stent graft migration, and iliac limb occlusion. They highlight that these complications cannot always be detected by visual inspection of computed tomography (CT) scans, and state that “an urgent need exists for more accurate and reliable postprocedure surveillance parameters by which to assess the behaviours of post-EVAR aneurysm sacs”. New technologies able to detect endoleaks and analyse their effects on the aneurysm sac are “highly relevant”, they specify.
The research group—which includes investigators from Bordeaux University Hospital and Imperial College London (London, UK)—note that EVAR surveillance relies on serial measurements of the maximal sac diameter despite significant inter- and intraobserver variability. However, Caradu et al write that volumetric measurements are more sensitive and that their implementation has been hampered by the time required for their implementation.
PRAEVAorta, which the authors describe as an “innovative, fully automated software,” had previously demonstrated “fast and robust detection of the characteristics of infrarenal [AAAs] on preoperative imaging studies”. In the present study, they note, the team assessed the robustness of these data on post-EVAR CT scans. Caradu and colleagues explain that they compared fully automatic and semiautomatic segmentation manually corrected by a senior surgeon using a dataset of 48 patients—48 early post-EVAR CT scans with 6,466 slices and 101 follow-up CT scans with 13,704 slices.
In JVS, the authors report that their analyses confirmed the excellent correlation of the post-EVAR volumes and surfaces and the proximal neck and maximum aneurysm diameters measured using the fully automatic and manually corrected segmentation methods (Pearson’s coefficient correlation >0.99; p<0.0001).
In addition, they note that a comparison between the fully automatic and manually corrected segmentation methods revealed a mean Dice similarity coefficient of 0.95±0.015, Jaccard index of 0.906±0.028, sensitivity of 0.929±0.028, specificity of 0.965±0.016, volumetric similarity of 0.973±0.018, and mean Hausdorff distance/slice of 8.7±10.8mm.
Finally, Caradu et al reveal that the segmentation time was nine times faster with the fully automatic method (2.5 minutes vs. 22 minutes per patient with the manually corrected method; p<0.0001), and that a preliminary analysis also demonstrated that a diameter increase of 2mm can actually represent a >5% volume increase.
The authors conclude that PRAEVAorta showed “excellent reproducibility” and “required less time and reduced the risk of human error”. They suggest that the software could become a “crucial adjunct” to endovascular aneurysm repair (EVAR) follow-up through early detection of sac evolution, which might reduce the risk of secondary rupture.
Caradu and colleagues acknowledged some limitations of their analysis in the discussion of their findings. The study was restricted to infrarenal AAAs, for example, due to the fact that the precise characterisation of patients with complex aneurysms, including the pararenal and visceral segments, “remains to be validated”. They write that further software development and validation are needed to enable automated reporting on the presence and type of endoleaks, adding that the development of volumetric analysis of the sealing zone “is also in progress”.
In closing, the investigators state that their results, in addition to current literature, “bring interesting perspectives to the use of AI for clinical practice”. In time, they believe, “these tools will be used for optimising preoperative planning and sizing, to better assess aneurysmal evolution, and to anticipate post-EVAR complications, with the potential to individually tailor postoperative surveillance protocols”.
In the spotlight: Application of AI in vascular diseases
Caradu et al’s work appeared in the September edition of JVS, which opened with a special communication on the application of AI in vascular surgery. In the article, Konstantinos Spanos (Larissa University Hospital, Larissa, Greece) and colleagues opine that AI and machine learning technologies are “rapidly revolutionising the healthcare system and patient management” and outline why vascular surgeons “should pay attention” to developments in the space.
“It is of paramount importance for healthcare providers and insurance companies to understand the state of AI technologies and how these can be used to improve the efficacy and safety of, and access to, healthcare services to achieve cost-effectiveness in healthcare status,” they write.
Spanos et al suggest that AI and its applications could benefit vascular physicians at both the clinical and the administrative levels. In regard to AAA disease, for example, they reference current international guidelines that highlight the importance of follow-up for these patients, something that the work of Caradu and colleagues addresses. “It is a clinical issue of paramount importance to predict which patients will require closer follow-up and potentially reintervention to prevent postoperative complications,” the authors stress.
Beyond assisting in the prediction of AAA growth and rupture risk, Spanos and colleagues hypothesise that the use of AI could also represent a method for determining the indications and optimal surgical treatment. They suggest that it might be able to assist in planning postoperative monitoring, and also in expediting the development of personalised decision-making and treatment for AAA patients, for instance.
Spanos et al also touch upon the applications of AI in cerebrovascular disease and peripheral arterial disease (PAD). For the former, the authors mention among various potential uses that AI could play a role in computer-aided prediction of stroke. They note that while the application of AI technologies to PAD is “in its infancy,” the potential is “tremendous”.
The authors also consider an economic perspective on the application of AI in healthcare. “Reported evidence is lacking regarding the economic effects of AI solutions for the healthcare system,” they write. They reference one meta-analysis on the topic, however they stress that “more comprehensive economic analyses are required before deciding for or against implementing AI technology in healthcare”.
Speaking to Vascular News, Caradu and senior author Eric Ducasse (Bordeaux University Hospital) remark on the the future direction of this the PRAEVAorta software: “In a research-use-only version, it integrates a prediction tool giving hints on how the aneurysm will evolve based on data classification of geometry and blood flow characteristics.
“Moreover, this AI is on the verge of proposing a consolidated solution for an advanced and complete 3D visualisation of the entire vascular network dedicated to vascular disease quantification (for instance in peripheral arterial disease) to enable early vascular disease diagnostic, follow-up, and prevention.”