Automated Aneurysm Boundary Detection and Volume Estimation Using Deep Learning

Figures

• 

  Figure 1. Example of the dataset. (A) 3D view front with annotation of the beginning (1) and end (2) of aneurysm. (B) cross-section view of the beginning of the aneurysm. (C) cross-section view of the end of aneurysm.

• 

  Figure 2. Vascular system segmentation with (A) UNet alone, versus (B) UNet + SAM2 on the same patient scan. The standalone UNet ceases to generate a mask in the aneurysmal region, visible between slices ~95 and ~125, whereas the UNet + SAM2 pipeline preserves a continuous aortic segmentation through the aneurysmal region.

• 

  Figure 3. Full aneurysm segmentation pipeline. (A) Original CTA scan; (B) SAM2-generated aorta mask overlaid on the scan after propagation; (C) Binary segmentation mask produced by SAM2; (D) Manual annotation showing normal aorta (red) and AAA region (blue). The sequence shows how the proposed approach extracts an aneurysm despite having no aneurysm labels during training.

• 

  Figure 4. Cross-sectional view of SAM2’s segmentation performance at the start (A) and end (B) boundaries of an aneurysm, where red shows the segmented region using UNet + SAM2 (refer to Figure 1 for anatomical reference).

• 

  Figure 5. Training and validation loss for LSTM.

• 

  Figure 6. USL performance on predicting the boundary of aneurysm compared to ground truth.

• 

  Figure 7. USL performance on predicting the aneurysm volume compared to ground truth.

• 

  Figure 8. USE performance on predicting the boundary of aneurysm compared to ground truth.

• 

  Figure 9. USE performance on predicting the aneurysm volume compared to ground truth.

• 

  Figure 10. Performance comparison between the USL and USE pipelines. (A) Bar chart of R2 for start boundary, end boundary, and surrogate volume, showing how much variance in the reference labels each model explains. (B,C) Box and whisker plots of absolute error for start and end slice indices; shorter bars indicate finer localization. (D) Bar charts of Mean Absolute Error (MAE) for start and end boundary for each approach. (E) Bar charts of Mean Squared Error (MSE) for start and end boundaries for each approach.

• 

  Figure 11. Violin plot of Dice distributions, revealing the full density of overlap scores for each pipeline. The solid curves depict the density distributions, while the horizontal dashed lines within each violin represent the quartiles of the data.

• 

  Figure 12. Scatter plot of predicted and ground truth surrogate volumes when aneurysm boundaries are selected randomly in the upper aorta, using the same window size as the ground truth. This illustrates the expected performance range in a worst-case, unguided boundary selection scenario.