On preserving anatomical detail in statistical shape analysis for clustering: focus on left atrial appendage morphology

Figures

• 

  FIGURE 1. (A) Location of the LAA on the left atria, with blood clot representation. (B) Visual display of two selected LAA cases, with anatomical nomenclature of ostium, neck, primary and secondary lobes and trabeculae. Note how these example LAA anatomies differ considerably in both shape and detail, which does not include the full breadth of LAA morphological variation.

• 

  FIGURE 2. (A) LAA mesh acquisition and processing prior to SSA and clustering. The upper far left shows an example slice of the CT image stack to achieve the lower left atrial segmentation. (B) The LAA position determined through a fully automatic detection algorithm (Martorana et al.). (C) Two examples of LAA point clouds before and after alignment through Super4PCS registration, followed by ICP & multiview registration of all possible pairs.

• 

  FIGURE 3. The simplified meshes (left to right) for two examples of LAA chicken wing and non-chicken wing morphologies from original trabeculated reconstruction, until full sample reduction. Note the visual loss in LAA trabeculae by 4-times sample reduction, and visual lobar definition loss by 8-times sample reduction. The data flow for the subsequent SSA and clustering methodology is also displayed in the bottom half of the figure–with SSA of all five datasets to determine SSA performance with greater sample reduction, followed by clustering comparison between the trabeculated dataset and one simplified dataset (4-times sample reduction).

• 

  FIGURE 4. SSA workflow in ShapeWorks. (A) Refers to the input dataset and applied SSA parameters. (B) Refers to the SSA process, which is multiscale in the initialisation and optimisation of particle placements, with increasing particles’ number (C) Refers to the outputs of the SSA (i.e., the PCA component scores after particle optimisation, the average shape and its variations) and the model performance evaluation metrics.

• 

  FIGURE 5. Shape variation captured by the first and second PC. In (A) the average shape with increasing reduction factor is presented. In (B) moving between 2 standard deviations on PC1 away from the average (±2σ) corresponds to chicken wing-like and non-chicken wing-like shape; with greater cumulative variance captured with increasing reduction with simpler shapes. In (C) moving between 2 standard deviations on PC2 away from the average (±2σ) corresponds more to secondary lobe size. Highlighted in blue are the two datasets (fully trabeculated and 4-times reduction) used for clustering comparisons.

• 

  FIGURE 6. Difference in SSA model evaluation scores compared to the trabeculated dataset with increasing number of PCs. As shown, increasing shape simplification (with increasing reduction factor) increases the amount of morphological variance captured at lower PCs (compactness), decreases the Euclidean distance between a sample shape and its closest training sample (specificity) and improves unseen shape representation (generalisation).

• 

  FIGURE 7. Dendrograms after hierarchical clustering of the trabeculated and simplified datasets. The dendrogram of trabeculated LAA morphologies indicated 23 as chicken wing-like, while the dendrogram of simplified morphology indicated 48. If categorised by a human expert, 21 LAAs are defined as chicken wing, suggesting that the trabeculated dendrogram is closer to human assessment.

• 

  FIGURE 8. The hierarchical cluster assignments are displayed on the trabeculated PCA distribution (PC1 on the horizontal axis against PC2 on the vertical axis), with AMI according to earlier clinical labels. The graph of the trabeculated dataset shows clear cluster separation between chicken wing (PC1 in the negative direction) and non-chicken wing cases, while the simplified dataset displays high overlap.