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
Noise suppression in photon-counting computed tomography using unsupervised Poisson flow generative models
KTH, School of Engineering Sciences (SCI), Physics, Particle Physics, Astrophysics and Medical Imaging. MedTechLabs, Karolinska University Hospital, 17164, Stockholm, Sweden.
Department of Clinical Neuroscience, Karolinska Institutet, 17164, Stockholm, Sweden; Department of Neuroradiology, Karolinska University Hospital, 17164, Stockholm, Sweden.
Department of Radiology, School of Medicine and Public Health, University of Wisconsin, 53705, Madison, WI, United States.
GE HealthCare, 53188, Waukesha, WI, United States.
Show others and affiliations
2024 (English)In: Visual Computing for Industry, Biomedicine, and Art, ISSN 2096-496X, Vol. 7, no 1, article id 24Article in journal (Refereed) Published
Abstract [en]

Deep learning (DL) has proven to be important for computed tomography (CT) image denoising. However, such models are usually trained under supervision, requiring paired data that may be difficult to obtain in practice. Diffusion models offer unsupervised means of solving a wide range of inverse problems via posterior sampling. In particular, using the estimated unconditional score function of the prior distribution, obtained via unsupervised learning, one can sample from the desired posterior via hijacking and regularization. However, due to the iterative solvers used, the number of function evaluations (NFE) required may be orders of magnitudes larger than for single-step samplers. In this paper, we present a novel image denoising technique for photon-counting CT by extending the unsupervised approach to inverse problem solving to the case of Poisson flow generative models (PFGM)++. By hijacking and regularizing the sampling process we obtain a single-step sampler, that is NFE = 1. Our proposed method incorporates posterior sampling using diffusion models as a special case. We demonstrate that the added robustness afforded by the PFGM++ framework yields significant performance gains. Our results indicate competitive performance compared to popular supervised, including state-of-the-art diffusion-style models with NFE = 1 (consistency models), unsupervised, and non-DL-based image denoising techniques, on clinical low-dose CT data and clinical images from a prototype photon-counting CT system developed by GE HealthCare.

Place, publisher, year, edition, pages
Springer Nature , 2024. Vol. 7, no 1, article id 24
Keywords [en]
Deep learning, Denoising, Diffusion models, Photon-counting CT, Poisson flow generative models
National Category
Computer graphics and computer vision Other Physics Topics
Identifiers
URN: urn:nbn:se:kth:diva-354278DOI: 10.1186/s42492-024-00175-6ISI: 001319529000001Scopus ID: 2-s2.0-85204916575OAI: oai:DiVA.org:kth-354278DiVA, id: diva2:1902936
Note

QC 20241008

Available from: 2024-10-02 Created: 2024-10-02 Last updated: 2025-05-08Bibliographically approved
In thesis
1. Deep learning approaches for denoising, artifact correction, and radiology report generation in CT and chest X-ray imaging
Open this publication in new window or tab >>Deep learning approaches for denoising, artifact correction, and radiology report generation in CT and chest X-ray imaging
2025 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

Medical imaging is a cornerstone of modern healthcare delivery, providing essential insights for effective diagnosis and treatment planning. Among the myriad imaging modalities, computed tomography (CT) and chest X-rays stand out for their widespread clinical use with approximately 400 million CT and 1.4 billion chest X-ray examinations are performed globally each year. Recent advancements in detector technology have given rise to photon-counting CT, which promises improved spatial and energy resolution along with enhanced low-dose imaging capabilities. However, elevated image noise and ring artifacts–stemming from higher spatial and energy resolution and inconsistencies in detector elements–pose significant hurdles, degrading image quality and complicating the diagnostic process. Beyond CT imaging, the volume of chest X-ray examinations continues to grow, placing increasing pressure on radiology departments that are already stretched thin. Moreover, advanced and innovate techniques in CT leads to a steady increase in the number of images that the radiologist are required to read, further exacerbating the workloads. To address these challenges, this thesis leverages generative artificial intelligence methods throughout the medical imaging value chain. For photon-counting CT imaging, this thesis address inverse problems using diffusion and Poisson flow models. Syn2Real synthesizes realistic ring artifacts to effciently generate training data for deep learning-based artifact correction. For image denoising, the thesis introduces methods that capitalize on the robustness of PFGM++ in supervised and unsupervised versions of posterior sampling Poisson flow generative models, and finally culminating in Poisson flow consistency models—a novel family of deep generative models that combines the robustness of PFGM++ with the effcient single-step sampling and the flexibility of consistency models. Moreover, this thesis works towards addressing the global shortage of radiologists, by improving medical vision-language models through CheXalign: a novel framework that leverages publicly available datasets, containing paired chest X-rays and radiology reports written in a clinical setting, and reference-based metrics to generate high quality preference data. This in turns enables the application of direct alignment algorithms that increase the probability of good reports, while decreasing the probability of bad ones, improving the overall results. Partial automation of chest X-ray radiology report generation—in which language models are used to draft initial reports—hold great promise for more effcient workflows, reducing burn-out, and allowing radiologists to allocate more time to more advanced imaging studies, such as photon-counting CT.

Abstract [sv]

Medicinsk avbildning är en hörnsten i den moderna sjukvården och ger avgörande insikter för e!ektiv diagnos och behandlingsplanering. Bland de många bildbehandlingsmetoderna utmärker sig datortomografi (CT) och lungröntgen för sin utbredda kliniska användning, där årligen cirka 400 miljoner CT-undersökningar och 1,4 miljarder lungröntgenundersökningar utförs globalt. Nya framsteg inom detektorteknik har lett till utvecklingen av fotonuppräknande CT, vilket lovar förbättrad rumslig och energiresolution samt förbättrade möjligheter för lågdosavbildning. Emellertid utgör förhöjt bildbrus och ringartifakter—till följd av högre rumslig och energiresolution samt inkonsekvenser i detektorelement—betydande hinder, vilket försämrar bildkvaliteten och komplicerar den diagnostiska processen. Utöver CT-avbildning fortsätter volymen av lungröntgenundersökningar att öka, vilket sätter ytterligare press på redan överbelastade radiologiavdelningar. Dessutom leder avancerade och innovativa tekniker inom CT till en stadig ökning av antalet bilder som radiologerna måste tolka, vilket ytterligare förvärrar arbetsbelastningen. För att möta dessa utmaningar utnyttjar denna avhandling generativa metoder inom artificiell intelligens genom hela värdekedjan för medicinsk avbildning. För fotonuppräknande CT-avbildning behandlar avhandlingen inversa problem med hjälp av diffusions- och Poisson-flödesmodeller. Syn2Real syntetiserar realistiska ringartifakter för att e!ektivt generera träningsdata för djupinlärnings-baserad artefaktkorrigering. För brusreducering i bilder introducerar avhandlingen metoder som utnyttjar robustheten hos PFGM++ i både övervakade och icke-övervakade versioner av posterior sampling Poisson-flödes generativa modeller, vilket kulminerar i Poisson-flödes konsistensmodeller—en ny familj av djupa generativa modeller som kombinerar robustheten hos PFGM++ med effektiv enkelsagsprovtagning och flexibiliteten hos konsistensmodeller. Dessutom arbetar denna avhandling för att tackla den globala bristen på radiologer genom att förbättra medicinska vision-språkmodeller med hjälp av CheXalign: ett nytt ramverk som utnyttjar o!entligt tillgängliga dataset, innehållande parade lungröntgenbilder och radiologiska rapporter skrivna i en klinisk miljö, samt referensbaserade mått för att generera högkvalitativ preferensdata. Detta möjliggör i sin tur tillämpningen av direkta justeringsalgoritmer som ökar sannolikheten för goda rapporter samtidigt som sannolikheten för dåliga minskar, vilket förbättrar de övergripande resultaten. Delvis automatisering av genereringen av lungröntgenrapporter—där språkmodeller används för att utarbeta initiala rapporter—lovar stora möjligheter till e!ektivare arbetsflöden, minskad utbrändhet och att radiologerna kan avsätta mer tid för mer avancerade avbildningsstudier, såsom fotonuppräknande CT.

Place, publisher, year, edition, pages
Universitetsservice US-AB, Sweden 2025, 2025
Series
TRITA-SCI-FOU ; 2025:29
Keywords
CT, photon-counting CT, chest X-rays, diffusion models, PFGM++, large language models, vision-language models, post-training, reinforcement learning from human feedback, direct alignment algorithms, CT, fotonräknande CT, lugnröntgen, diffusionsmodeller, PFGM++, stora språkmodeller, vision-språkmodeller, efterträning, förstärkningsinlärning från mänsklig feedback, direktjusteringsalgoritmer
National Category
Radiology and Medical Imaging Other Physics Topics
Research subject
Physics, Biological and Biomedical Physics
Identifiers
urn:nbn:se:kth:diva-363233 (URN)978-91-8106-316-5 (ISBN)
Public defence
2025-06-05, FD5, Roslagstullsbacken 21, Stockholm, 09:15 (English)
Opponent
Supervisors
Note

QC 2025-05-09

Available from: 2025-05-09 Created: 2025-05-08 Last updated: 2025-07-01Bibliographically approved

Open Access in DiVA

No full text in DiVA

Other links

Publisher's full textScopus

Authority records

Hein, DennisDanielsson, MatsPersson, Mats

Search in DiVA

By author/editor
Hein, DennisDanielsson, MatsPersson, Mats
By organisation
Particle Physics, Astrophysics and Medical Imaging
Computer graphics and computer visionOther Physics Topics

Search outside of DiVA

GoogleGoogle Scholar

doi
urn-nbn

Altmetric score

doi
urn-nbn
Total: 74 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