Artificial intelligence for fully automatic segmentation of abdominal aortic aneurysm using convolutional neural networks

Abstract

We were pleased to read that our publication on implementing artificial intelligence to provide a fully automatic segmentation of abdominal aortic aneurysm (AAA) from computed tomography (CT) angiography1Caradu C. Spampinato B. Vrancianu A.M. Bérard X. Ducasse E. Fully automatic volume segmentation of infrarenal abdominal aortic aneurysm computed tomography images with deep learning approaches versus physician controlled manual segmentation.J Vasc Surg. 2021; 74: 246-256Abstract Full Text Full Text PDF Scopus (6) Google Scholar had aroused the interest of teams working on similar approaches. We agree with Lareyre et al2Lareyre F. Adam C. Carrier M. Dommerc C. Mialhe C. Raffort J. A fully automated pipeline for mining abdominal aortic aneurysm using image segmentation.Sci Rep. 2019; 9: 13750Crossref PubMed Scopus (21) Google Scholar that imaging is a key step in the management of AAAs, whether preoperatively (providing a threshold of size/volume and sizing for endovascular aneurysm repair) or for postoperative follow-up. This is especially true in complex anatomies with short necks and angulations. Although we could not provide a complete description of the pipeline, algorithms, and formulas for patent purposes, we can provide the main sequential steps. The PRAEVAorta software implements deep learning models, which are an assembly of neural networks known as convolutional neural networks with five layers for approximately 60 million parameters. Presently, it uses two deep learning models, one for the lumen and another for the thrombus. Each model is coupled with a "confirmation" algorithm based on morphological "closing" and "opening" operations as well as on the analysis of pixel intensities. The CT scans came from a wide range of peripheral centers' referrals, and we observed no performance difference between four different scan manufacturers. The dataset was restricted to infrarenal AAA (excluding pararenal, thoracoabdominal, ruptured aneurysms, and aortic dissections) to standardize the algorithms and check their validity as a first step before extending it to more difficult anatomies. It now seems to work just as well on post-endovascular aneurysm repair CT scans, despite the artifacts related to stents and radiopaque markers, noncontrast CT scans, aortic dissections, carotid arteries, and lower limbs. We wish to increase each of these series before publishing those results. The next step will be to include the paravisceral aorta, but we still need time to retrieve enough CT scans, feed them to the machine, compare them with a semiautomatic method for validation, and analyze the results. As others,3Raffort J. Adam C. Carrier M. Ballaith A. Coscas R. Jean-Baptiste E. et al.Artificial intelligence in abdominal aortic aneurysm.J Vasc Surg. 2020; 72: 321-333.e1Abstract Full Text Full Text PDF PubMed Scopus (27) Google Scholar,4Dey D. Slomka P.J. Leeson P. Comaniciu D. Shrestha S. Sengupta P.P. et al.Artificial intelligence in cardiovascular imaging: JACC state-of-the-art review.J Am Coll Cardiol. 2019; 73: 1317-1335Crossref PubMed Scopus (150) Google Scholar we are convinced that in a near future, these approaches could become an incredible therapeutic aid decision-making in clinical practice for diagnosis, planning of surgical repair, and follow-up. We would like to thank Romain Leguay and Florian Bernard (Nurea) for graciously providing the software PRAEVAorta.