kth.sePublications KTH
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
3D Helical CT Reconstruction With a Memory Efficient Learned Primal-Dual Architecture
KTH, School of Engineering Sciences (SCI), Mathematics (Dept.), Mathematics (Div.).ORCID iD: 0000-0002-6648-2378
Univ Novi Sad, Fac Tech Sci, Novi Sad 21000, Serbia..
Philips Innovat Technol, D-22335 Hamburg, Germany..
KTH, School of Engineering Sciences (SCI), Mathematics (Dept.).ORCID iD: 0000-0002-1118-6483
2024 (English)In: IEEE Transactions on Computational Imaging, ISSN 2573-0436, E-ISSN 2333-9403, Vol. 10, p. 1414-1424Article in journal (Refereed) Published
Abstract [en]

Deep learning based computed tomography (CT) reconstruction has demonstrated outstanding performance on simulated 2D low-dose CT data. This applies in particular to domain adapted neural networks, which incorporate a handcrafted physics model for CT imaging. Empirical evidence shows that employing such architectures reduces the demand for training data and improves upon generalization. However, their training requires large computational resources that quickly become prohibitive in 3D helical CT, which is the most common acquisition geometry used for medical imaging. This paper modifies a domain adapted neural network architecture, the Learned Primal-Dual (LPD), so that it can be trained and applied to reconstruction in this setting. The main challenge is to reduce the GPU memory requirements during the training, while keeping the computational time within practical limits. Furthermore, clinical data also comes with other challenges not accounted for in simulations, like errors in flux measurement, resolution mismatch and, most importantly, the absence of the real ground truth. To the best of our knowledge, this work is the first to apply an unrolled deep learning architecture for reconstruction on full-sized clinical data, like those in the Low dose CT image and projection data set (LDCT).

Place, publisher, year, edition, pages
Institute of Electrical and Electronics Engineers (IEEE) , 2024. Vol. 10, p. 1414-1424
Keywords [en]
Computed tomography, Image reconstruction, Computer architecture, Training, Three-dimensional displays, Neural networks, Reconstruction algorithms, deep learning, helical acquisition, primal-dual, clinical data
National Category
Computer graphics and computer vision
Identifiers
URN: urn:nbn:se:kth:diva-355183DOI: 10.1109/TCI.2024.3463485ISI: 001327401000001Scopus ID: 2-s2.0-85204642534OAI: oai:DiVA.org:kth-355183DiVA, id: diva2:1907965
Note

QC 20241024

Available from: 2024-10-24 Created: 2024-10-24 Last updated: 2025-02-19Bibliographically approved
In thesis
1. Data-driven Image Reconstruction in Computed Tomography
Open this publication in new window or tab >>Data-driven Image Reconstruction in Computed Tomography
2025 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

This thesis focuses mainly on improving the quality of reconstruction in computed tomography (CT), which is an imaging techniquethat aims to reconstruct the interior of an object from a set of X-ray projections obtained from different viewpoints. Mathematically, this is an inverse problem that is often ill-posed and, therefore, requires some sort of regularization. Recently, this research field has been dominated by data-driven approaches, and this thesis is not an exception. In contrast to model-based methods that assume that a reconstructed object possesses certain predefined properties, data-driven methods use the statistical information obtained from a set of similar objects to improve reconstruction of a new object from the same class. The recent development of computer hardware has allowed us to extract and store this statistical information from a large set of samples, and thereforeis responsible for the increased popularity of these methods. However, the problem with these approaches is that the most efficient of them, such as deep neural networks, lack interpretability, and their extraordinary empirical performance is not fully justified from the theoretical perspective. Another big challenge concerning CT in particular is to move from methods that perform very well on toy experiments conducted on simulated low-dimensional data to methods that could be used in real-life applications, such as medical imaging.

This thesis explores several directions that could potentially address the above issues. The first is data-driven optimization that can be used to reduce the number of iterations needed to obtain the final reconstruction when it is defined as a solution to an optimization problem. Such optimization problems appear within a classic regularization framework.

Next, we revisit dictionary learning, which can be seen as a predecessor to data-driven methods that go under the caption “deep learn-ing”. The advantage of dictionary learning is in its relative mathematical simplicity and interpretability. We see this as a bridge between well-understood but somewhat limited model-based approaches and a black-box paradigm of deep learning. Since reconstruction using learned dictionaries is defined as an optimization problem, data-driven optimization comes in useful here as well.

Finally, this thesis addresses the problem of upscaling state-of-the-art deep learning architectures so that they can be applied to clinical CT data. We show that certain modifications in the architecture combined with engineering techniques allow us to do that without relyingon super-computing resources.

Abstract [sv]

Avhandlingen fokuserar huvudsakligen på att förbättra rekonstruktionskvaliteten inom datortomografi (DT), vilket är en  avbildningsteknik som syftar till att avbilda insidan av ett objekt från en uppsättning genomlysningsbilder som fås genom att belysa objektet med röntgen från olika riktningar. Matematiskt sett svarar detta mot ett illaställt inversproblem. För att hantera illaställdheten krävs regularisering. På senare tid har detta forskningsområde dominerats av data-drivna metoder, och denna avhandling är inget undantag. Till skillnad från modellbaserade metoder, som antar att objektet som avbildas har vissa fördefinierade egenskaper, bygger data-drivna metoder på statistisk information erhållen från en uppsättning liknande objekt för att förbättra rekonstruktionen av ett nytt objekt från samma klass. Den senaste utvecklingen inom datorhårdvara har gjort det möjligt att extrahera och lagra sådan statistisk information från stora mängder data och är därmed en av anledningarna till dessa metoders ökade popularitet. Problemet med dessa metoder är dock att de mest effektiva, såsom djupa neurala nätverk, saknar tolkbarhet, och deras extraordinära empiriska prestanda är inte fullt ut motiverad ur ett teoretiskt perspektiv. En annan stor utmaning, särskilt inom DT, är att övergå från metoder som fungerar mycket bra i teoretiska experiment med simulerade lågdimensionella data, till metoder som kan användas i praktiska tillämpningar, såsom medicinsk bilddiagnostik.

Avhandlingen utforskar flera riktningar som potentiellt kan adressera ovanstående problem. Den första är data-drivna optimeringsmetoder som kan användas för att minska antalet iterationer som behövs för att uppnå en slutgiltig rekonstruktion, när denna definieras som en lösning till ett optimeringsproblem. Sådana optimeringsproblem förekommer inom klassiska regulariseringsramverk.

Nästa område som undersöks är ordbokslärande (dictionary learning), vilket kan ses som en föregångare till data-drivna metoder under rubriken ``djupinlärning''. Fördelen med ordbokslärande ligger i dess relativa matematiska enkelhet och tolkbarhet. Detta ses som en brygga mellan välförstådda men något begränsade modellbaserade metoder och den ``svart låda''-paradigm som djupinlärning representerar. Eftersom rekonstruktion med hjälp av inlärda ordböcker definieras som ett optimeringsproblem, blir data-drivna optimeringsmetoder också användbara här.

Slutligen adresserar avhandlingen problemet med att skala upp moderna djupinlärningsarkitekturer så att de kan tillämpas på kliniska DT-data. Vi visar att vissa modifieringar i arkitekturen, kombinerat med ingenjörstekniker, möjliggör detta utan att behöva förlita sig på superdatorresurser.

Place, publisher, year, edition, pages
Stockholm: KTH Royal Institute of Technology, 2025
Series
TRITA-SCI-FOU ; 2025:11
National Category
Medical Imaging
Research subject
Applied and Computational Mathematics
Identifiers
urn:nbn:se:kth:diva-360160 (URN)978-91-8106-218-2 (ISBN)
Public defence
2025-03-17, Lecture Hall F3, Lindstedtsvägen 22, Stockholm, 13:00 (English)
Opponent
Supervisors
Funder
Swedish Foundation for Strategic Research, AM13-0049Vinnova, 2015-06759
Note

QC 2025-02-19

Available from: 2025-02-19 Created: 2025-02-19 Last updated: 2025-02-25Bibliographically approved

Open Access in DiVA

No full text in DiVA

Other links

Publisher's full textScopus

Authority records

Rudzusika, JevgenijaÖktem, Ozan

Search in DiVA

By author/editor
Rudzusika, JevgenijaÖktem, Ozan
By organisation
Mathematics (Div.)Mathematics (Dept.)
In the same journal
IEEE Transactions on Computational Imaging
Computer graphics and computer vision

Search outside of DiVA

GoogleGoogle Scholar

doi
urn-nbn

Altmetric score

doi
urn-nbn
Total: 391 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf