Study design

A prospective single-centre, observational proof-of-concept study.

Patients

Fifty patients from the Copenhagen Aortic CoHort (COACH) were included. COACH is a prospective cohort of patients with AAAs with diameters above 30 mm and monitored with research-orientated US surveillance [ ]. Enrolment in COACH relies on research staff and equipment availability. Exclusion criteria were aorto-iliac aneurysms; suprarenal aortic aneurysms; prior aortic or iliac artery repair; non-compliance with surveillance; or inability to provide informed consent. US acquisitions were systematically collected and stored.

US image acquisition

Two US operators independently and non-consecutively recorded two 10-s cross-sectional 2D-US time-resolved B-mode sequences (video clips called “cineloops”) of the AAA at the maximal AP-diameter during breath holds using an identical US system (Epiq Elite US system and C5-1 transducer, Philips Medical Systems, Bothell, WA, USA). Thus, for each patient, four cineloops were planned (two recordings × two operators). Acquisitions were subsequently transferred to a core lab, where two readers (USL and MIB) mutually and independently conducted the image post-processing in a randomised order via randomised.org, using a prototype software (Philips Health Technology innovation). The image quality of every cineloop was evaluated and graded by a clinical expert reader (USL). Cineloops of poor image quality ( n = 9, 4.5%) were not used for further analysis. A few of the planned acquisitions were either not conducted or failed to upload to our database ( n = 7, 3.5%).

Strain

Strains are the percentage deformation of the AAA vessel wall when force is applied to it, for example, by the blood pressure during the heart cycle. When overlaying a grid on the AAA vessel wall, maximum principal strain is the largest deformation (or fibre stretch) value in any direction (called the first “principal direction”). Higher values indicate softer tissues and lower values indicate stiffer tissues. Principal strain can then be projected onto two main directions of movement along the AAA vessel wall: circumferential (stretch or compression) or radial (thickening or thinning); see Figure A1 .

Image post-processing

The image post-processing is described in detail in previous work [ ], with the notable difference that the present study includes principal strain in addition to circumferential and radial strain. The process can be summarised as a 3-step procedure: (i) manual segmentation and automatic tracking, (ii) automatic strain map generation and (iii) strain pattern extraction.

• 1)
  

  Manual segmentation and automatic tracking

  
  

  Tracking points were manually positioned at the initial frame of the cineloop. This was done circumferentially at the interface between the aneurysm wall and lumen ( Fig. 1 b). Subsequently, an automated speckle-tracking process was employed based on image gradients within an automatic region of interest ( Fig. 1 c).

  
  
• 2)
  

  Automatic strain map generation

  
  

  A deformation grid was computed based on the tracking procedure, and from this, strain over the cardiac cycle—from end-diastole to peak systole—was directly calculated. The resulting strain data were superimposed on the cineloop. This generated a comprehensive strain map ( Fig. 1 d) illustrating different zones of AAA strain, expressed as percentage changes over time. Results from all cardiac cycles were then averaged.

  
  
• 3)
  

  Strain pattern extraction and normalization

  
  

  The obtained strain maps contain local information along an irregular vessel wall. To enable inter-case comparisons of strain maps, strain values were radially averaged for each degree of the AAA circumference with 0° defined as the bottom of the strain map (toward the spine), see Figure 1 e. Thus, the total information of an AAA strain map is summarised as a polar plot, henceforth designated the “strain pattern,” see Figure 1 f (where the black circle represents the reference null strain). Strain patterns were then smoothed to reduce sensitivity to noise.

The whole post-processing phase spans 3–5 min, in off-cart software. With future improvement the software could be operationally on-cart version for full clinical integration.

Functional data analysis of strain patterns

To analyse strain patterns without reducing them to single parameter characterisations, functional data analysis (FDA) with alignment was performed [ ]. Alignment is the process of elastically synchronizing (or “warping”) functional data sets to a common reference, enabling comparative analysis and inference across patients. The mean strain pattern was calculated on smoothed and elastically aligned strain patterns to reduce noise and account for minor scan variations (e.g., in insonation angles). Further, using function principal component analysis (FPCA), the strain patterns were decomposed into their principal modes of variations (PMV) with associated scores for each cineloop. The scores for each PMV was analysed for correlation with AAA size using Spearman’s rank correlation. A positive or negative coefficient of 0–0.3 was considered weak correlation, 0.3–0.5 low, 0.5–0.7 moderate, 0.7–0.9 high and 0.9–1.0 very high [ ]. p values were corrected using the Benjamini–Hochberg false discovery rate (FDR) procedure. Only FDR- p values <0.05 were considered statistically significant.

Clustering

Subsequently, spectral consensus clustering [ , ] was then used to split the cineloops into an optimal number of clusters based on the PMV scores.

Single parameter strain characterisations

As described in other work on strain mapping [ ], single numeric parameter strain characterisations can be calculated from the end-diastolic to peak-systolic strain based on the principal strains across the vessel wall: mean strain; pulse pressure-normalised mean strain (PPPS); peak strain; heterogeneity index, that is, the ratio of the SD of wall strain to the mean wall strain; and local strain ratio, that is, the ratio between the peak strain and mean strain.

Single parameter strain characteristics were summarised across all patients with frequency-weighted medians and inter-quartile ranges (IQR), due to non-normality. All single parameter strain characteristics were analysed for correlation with AAA size using Spearman’s rank correlation similarly to the PMVs.

Reproducibility

For strain pattern clusters, the “cluster reproducibility” was calculated, which was defined as the percentage of an individual patient’s four cineloops that belong to the majority (or tie for majority) cluster (e.g., three out of four scans within the same cluster equals 75%). By aligning all of a patient’s scans to each other (within-patient alignment), amplitude and phase variations for each patient could be extracted. These were averaged across patients for an overall amplitude-phase variation ratio.

For single parameter characterisations, Bland–Altman limits of agreement (LoA) on inter-reader comparisons were calculated, comparing readers USL and MIB on the same cineloops. The LoA were quantile-based, due to non-normality, and were converted into a range of variance (RoV, i.e., half the quantile distance) and displayed as a percentage of the median. To calculate LoA for inter-operator considerations, a linear mixed-effects model was utilised, accounting for multiple acquisitions by the same operator.

All statistical analyses described above were performed in R version 4.3.3 using RStudio version 2023.12.1 build 402 (Posit software, PBC, Boston MA, USA). The packages: Hmisc, mgcv, fdasrvf, fdapace and M3C were used for weighted statistics, strain pattern smoothing, alignment, FPCA and clustering, respectively.