Salvi, Anish
(2023)
Machine Learning for Abdominal Aortic Aneurysm Characterization from Standard-Of-Care Computed Tomography Angiography Images.
Master's Thesis, University of Pittsburgh.
(Unpublished)
This is the latest version of this item.
Abstract
Abdominal aortic aneurysms (AAAs) are dilations in the descending aorta which can result in internal bleeding when ruptured, leading to hospitalization or death. AAAs are commonly asymptomatic and discovered by happenstance during imaging tests, including computed tomography (CT) and its blood vessel enhancing counterpart computed tomography angiography (CTA). However, a past evaluation indicates that radiologists correctly identified and referred to monitoring only 32%, or 43 of 133 AAAs, from 3292 CTs. AAAs with larger diameters (> 5 cm) are recommended for elective repair; however, < 5 cm AAAs may have a rupture rate as high as 23%. Utilizing diameter as a one size fits all approach fails to consider intraluminal thrombus (ILT) and calcifications, clinically relevant attributes associated with elevated rupture risk. While a prior study indicated that type I and III endoleaks, linked with incorrect graft positioning during elective repair, had an incidence of only 6.4%, these complications require urgent medical attention. Surgical planning may benefit from greater understanding of AAA geometry. There remains a critical need for the automated discovery, visualization, and elective repair indication of AAAs. Having explored the novel field at the intersection of state-of-the-art machine learning and standard-of-care medical imaging, we develop deep learning models to aid the detection, segmentation, and classification of AAAs based on pre-operative CTA characteristics while observing the frame-of-reference. We describe computational methods which include bounding box localization as a precursor to high-resolution segmentation, patch-based segmentation of medical image sub-volumes, image transformers that identify AAA severity, and a vision transformer that provides heatmaps indicative of AAA severity prediction. We find that 1) our memory-efficient bounding box method outperforms conventional neural network based AAA lumen segmentation, 2) patch-based AAA wall segmentation has improved performance as compared to our memory efficient computational pipeline for asymptomatic cases, 3) image transformers approach and even beat the accuracy achieved by rudimentary classifiers (i.e., differentiating between asymptomatic v. symptomatic AAAs) when leveraging embeddings derived from class specific segmentation models, and 4) vision transformers not only predict AAA severity accurately, but localize the disease by its anatomical basis. In sum, we make key contributions to scientific literature concerning medical imaging and machine learning through our computational methods of AAA interpretation.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
19 January 2023 |
| Date Type: |
Publication |
| Defense Date: |
7 November 2022 |
| Approval Date: |
19 January 2023 |
| Submission Date: |
7 November 2022 |
| Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
| Number of Pages: |
91 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
Swanson School of Engineering > Bioengineering |
| Degree: |
MS - Master of Science |
| Thesis Type: |
Master's Thesis |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
abdominal aortic aneurysms, deep learning, convolutional neural networks, vision transformers, image transformers |
| Date Deposited: |
19 Jan 2023 19:26 |
| Last Modified: |
19 Jan 2023 19:26 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/43853 |
Available Versions of this Item
-
Machine Learning for Abdominal Aortic Aneurysm Characterization from Standard-Of-Care Computed Tomography Angiography Images. (deposited 19 Jan 2023 19:26)
[Currently Displayed]
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}