Semi-supervised Virtual Regression of Aortic Dissections Using 3D Generative Inpainting

Antonio Pepe*, Gabriel Mistelbauer, Christina Schwarz-Gsaxner, Jianning Li, D. Fleischmann, Dieter Schmalstieg, Jan Egger*

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Aortic dissection (AD) is a condition of the main artery of the human body, resulting in the formation of a new flow channel, or false lumen (FL). The disease is usually diagnosed with a computed tomography angiography (CTA) scan during the acute phase. A better understanding of the causes of AD requires knowledge of aortic geometry prior to the event, which is available only in very rare circumstances. In this work, we propose an approach to reconstruct the aorta before the formation of a dissection by performing 3D inpainting with a two-stage generative adversarial network (GAN). In the first stage of our two-stage GAN, a network is trained on the 3D edge information of the healthy aorta to reconstruct the aortic wall. The second stage infers the image information of the aorta to reconstruct the entire dataset. We train our two-stage GAN with 3D patches from 55 non-dissected aortic datasets and evaluate it on 20 more non-dissected datasets, demonstrating that our proposed 3D architecture outperforms its 2D counterpart. To obtain pre-dissection aortae, we mask the entire FL in AD datasets. Finally, we provide qualitative feedback from a renown expert on the obtained pre-dissection cases
Original languageEnglish
Title of host publicationThoracic Image Analysis - Second International Workshop, TIA 2020, Held in Conjunction with MICCAI 2020, Proceedings
Subtitle of host publicationSecond International Workshop, TIA 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 8, 2020, Proceedings
EditorsJens Petersen, Raúl San José Estépar, Alexander Schmidt-Richberg, Sarah Gerard, Bianca Lassen-Schmidt, Colin Jacobs, Reinhard Beichel, Kensaku Mori
Place of PublicationCham
PublisherSpringer
Pages130-140
Number of pages11
ISBN (Electronic)978-3-030-62469-9
ISBN (Print)978-3-030-62468-2
DOIs
Publication statusE-pub ahead of print - Nov 2020
Event2nd International Workshop on Thoracic Image Analysis - Virtuell, Peru
Duration: 8 Oct 20208 Oct 2020

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12502 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference2nd International Workshop on Thoracic Image Analysis
Abbreviated titleTIA 2020
CountryPeru
CityVirtuell
Period8/10/208/10/20

Keywords

  • Aortic dissection
  • Deep learning
  • Edge reconstruction
  • Generative adversarial networks
  • Inpainting

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint Dive into the research topics of 'Semi-supervised Virtual Regression of Aortic Dissections Using 3D Generative Inpainting'. Together they form a unique fingerprint.

Cite this