Fully automated assessment of post-TEVAR follow-up CT scan using Deep Learning-Based Segmentation.
Authors
Affiliations (6)
Affiliations (6)
- Vascular Artificial Intelligence Laboratory (VAI-Lab), Department of Integrated Surgical and Diagnostic Sciences (DISC), University of Genoa, Genoa, Italy; Vascular and Endovascular Surgery Unit, IRCCS Ospedale Policlinico San Martino, 16132 Genoa, Italy. Electronic address: [email protected].
- Vascular Artificial Intelligence Laboratory (VAI-Lab), Department of Integrated Surgical and Diagnostic Sciences (DISC), University of Genoa, Genoa, Italy; Department of Experimental Medicine (DIMES), University of Genoa, Genoa, Italy.
- Camelot Biomedical Systems, Genoa, Italy.
- Vascular and Endovascular Surgery Unit, IRCCS Ospedale Policlinico San Martino, 16132 Genoa, Italy; Department of Integrated Surgical and Diagnostic Sciences (DISC), University of Genoa, 16132 Genoa, Italy.
- Vascular Artificial Intelligence Laboratory (VAI-Lab), Department of Integrated Surgical and Diagnostic Sciences (DISC), University of Genoa, Genoa, Italy.
- Department of Civil Engineering and Architecture, University of Pavia, 27100 Pavia, Italy; 3D and Computer Simulation Laboratory, IRCCS Policlinico San Donato, 20097 San Donato Milanese, Italy.
Abstract
The objective of this study was to apply an artificial intelligence (AI) pipeline for the automatic analysis of follow-up Computed Tomography Angiography (CTA) after Thoracic Endovascular Aortic Repair (TEVAR). A deep-learning network was developed to automatically measure the mean diameters of the proximal (D_LZP) and distal (D_LZD) landing zones, stent length (L), maximum aneurysm diameter (D_MAX), and aneurysm volume. Segmentation accuracy was assessed with the Dice Similarity Coefficient (DSC), and agreement with manual measurements using the Intraclass Correlation Coefficient (ICC). Manual measurements were obtained by an experienced vascular surgeon using dedicated software (EndoSize), based on standardized centerline landmarks. The study included 45 TEVAR patients; 3 Computed Tomography (CT) scans (6.6%) were excluded due to segmentation failure, leaving 84 CT scans for analysis (42 preoperative and 42 follow-up). At 1 month and 1 year, D_LZP was 32 ±4.7 mm (ICC 0.86) and 33.41 ±5.8 mm (ICC 0.78), D_LZD 30.15 ±4.54 mm (ICC 0.97) and 31.35 ±4.85 mm (ICC 0.83), while stent length remained stable (∼212 mm, ICC >0.98). D_MAX decreased from 57.4 ±14.8 mm to 55.5 ±13.2 mm (p<0.0001), and aneurysm volume from 68.6 ±86.1 mm<sup>3</sup> to 54.3 ±103.7 mm<sup>3</sup> (p<0.0001), with a strong correlation between changes in D_MAX and volume (r=0.77, p<0.0001). The proposed AI-based pipeline enables reliable and reproducible quantification of stent-graft landing zones after TEVAR. The strong agreement with manual measurements and the detection of significant morphological changes over time support its potential for standardized, objective, and clinically relevant follow-up.