University of Twente Student Theses
Novel Deep Learning Methods for Modeling and Prediction of Abdominal Aortic Aneurysm Growth
Hofman, M. (2023) Novel Deep Learning Methods for Modeling and Prediction of Abdominal Aortic Aneurysm Growth.
|
PDF
31MB |
| Abstract: | Abdominal aortic aneurysms (AAAs) are focal dilatations of the abdominal aorta that, if left untreated, can rupture with high mortality rate. In this work, we present novel deep learning methods for modeling and prediction of local AAA growth. Using implicit neural representations (INRs), we obtained personalized, continuous representations of AAA shapes evolving over time, based on highly sparse and irregularly spaced longitudinal image data. We represent the AAA’s outer wall evolving over time as the zero level set of its signed distance function (SDF), whichweembedinamultilayerperceptron(MLP)thatoperatesonspaceandtime. Weoptimize this INR using automatically extracted AAA segmentations in longitudinal CTA data. This network is conditioned on spatio-temporal coordinates and therefore represents the evolving AAA shape at any spatial resolution and any point in time. Using regularization on spatial and temporal gradients of the SDF, we observe that our model can accurately interpolate AAA shapes evolving over time, with average surface distances (ASDs) ranging from 0.627 to 4.443 mm. This personalized approach for modeling AAA evolution, however, does not generalize easily to new AAA patients, limiting its adoption in clinical practice. To address this, we propose a graph convolutional network (GCN) for prediction of local AAA growth, operating on surface mesh representations of the AAA’s outer wall. We optimize this network using continuous representations of evolving AAA shapes from multiple patients, that we obtained using the INRs. By conditioning the GCN model on a time step, we can predict AAA growth over any desired future time point. We demonstrate the GCN’s performance to predict AAA shapes with diameter profiles along the AAA centerlines. The results indicate that our model can predict local AAA growth in the right direction specifically in the dilated part of the aorta, leaving the healthy parts unaffected. Our proposed pipeline, including automatic segmentation, continuous AAA surface representation, and predicting local AAA growth on the surface, holds potential clinical value for more personalized, pro-active AAA surveillance. |
| Item Type: | Essay (Master) |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 31 mathematics, 44 medicine, 50 technical science in general, 54 computer science |
| Programme: | Biomedical Engineering MSc (66226) |
| Link to this item: | https://purl.utwente.nl/essays/96640 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page