Vis enkel innførsel

dc.contributor.advisorRobert, Jenssen
dc.contributor.authorWickstrøm, Kristoffer Knutsen
dc.date.accessioned2022-10-13T22:29:14Z
dc.date.available2022-10-13T22:29:14Z
dc.date.embargoEndDate2027-10-28
dc.date.issued2022-10-28
dc.description.abstractRetten til helse er en grunnleggende menneskerettighet, men mange utfordringer står overfor de som ønsker å etterleve denne retten. Mangel på utdannet helsepersonell, økte kostnader og en aldrende befolkning er bare noen få eksempler på nåværende hindringer i helsesektoren. Å takle slike problemer er avgjørende for å gi pålitelig helsehjelp med høy kvalitet til mennesker over hele verden. Mange forskere og helsepersonell mener at et datadrevet helsevesen har potensial til å løse mange av disse problemene. Datadrevne metoder er basert på algoritmer som lærer å utføre oppgaver ved å identifisere mønstre i data, og forbedres ofte i takt med at mer data blir samlet inn. En sentral drivkraft i moderne datadrevet helsevesen er dyp læring. Dyp læring er en del av representasjonslæringsfeltet, hvor målet er å lære en datarepresentasjon som er gunstig for å utføre en gitt oppgave. Dyp læring har ført til store forbedringer i viktige helsedomener som bilde og språkbehandling. Imidlertid mangler dyplæringsalgoritmer tolkbarhet, gir ikke uttryk for usikkerhet, og har vanskeligheter når de får i oppgave å lære fra data uten menneskelige annoteringer. Dette er grunnleggende begrensninger som må adresseres for at et datadrevet helsevesen basert på dyp læring skal nå sitt fulle potensial. For å takle disse begrensningene foreslår vi ny metodikk innen dyp læring. Vi presenterer de første metodene for å fange opp usikkerhet i forklaringer av prediksjoner, og vi introduserer det første rammeverket for å forklare representasjoner av data. Vi introduserer også en ny metode som utnytter domenekunnskap til å trekke ut klinisk relevante attributer fra medisinske bilder. Vårt fokus er på helseapplikasjoner, men den foreslåtte metodikken kan brukes i andre domener også. Vi tror at innovasjonene i denne oppgaven kan spille en viktig rolle i å skape pålitelige dyplæringsalgoritmer som kan lære av umerkede data.en_US
dc.description.abstractThe right to health is a fundamental human right, but numerous challenges face those who wish to comply. Shortage of trained health personnel, increases in costs, and an aging population are just a few examples of obstacles that arise in the healthcare sector. Tackling such problems is crucial to provide high quality and reliable healthcare to people around the world. Many researchers and healthcare professionals believe that data-driven healthcare has the potential to solve many of of these problems. Data-driven methods are based on algorithms that learn to perform tasks by identifying patterns in data, and often improve in line with the amount of data. A key driving force in contemporary data-driven healthcare is deep learning, which is part of the representation learning field where the goal is to learn a data representation that is beneficial for performing some task. Deep learning has lead to major improvements in important healthcare domains such as computer vision and natural language processing. However, deep learning algorithms lack explainability, do not provide a notion of uncertainty, and struggle when tasked with learning from unlabeled data. These are fundamental limitations that must be tackled for deep learning-based data-driven healthcare to reach its full potential. Towards tackling these limitation, we propose new methodology within the field of deep learning. We present the first methods for capturing uncertainty in explanations of predictions, and we introduce the first framework for explaining representations of data. We also introduce a new method that utilizes domain knowledge to extract clinically relevant features from medical images. While our emphasis is on healthcare applications, the proposed methodology can be employed in other domains as well, and we believe that the innovations in this thesis can play an important part in creating trustworthy deep learning algorithms that can learn from unlabeled data.en_US
dc.description.doctoraltypeph.d.en_US
dc.description.popularabstractThe right to health is a fundamental human right, but numerous challenges face those who wish to comply. Shortage of trained health personnel, increases in costs, and an aging population are just a few examples of obstacles that arise in the healthcare sector. Tackling such problems is crucial to provide high quality and reliable healthcare to people around the world. Recent advances in AI has the potential to revolutionize healthcare by learning to perform tasks from large amounts of data. However, current solutions suffer from some fundamental limitations; they lack trustworthiness and struggle to learn without human guidance. This thesis presents new methods that address these limitations. We develop AI systems that can explain their decision and indicate their degree of confidence, and also extract useful information from data without human intervention. The methods introduced in this thesis could play an important role in ushering the next generation of healthcare.en_US
dc.description.sponsorshipThis thesis was financially supported by the UiT Thematic Initiative “Data-Driven Health Technology”.en_US
dc.identifier.isbn978-82-8236-501-7
dc.identifier.urihttps://hdl.handle.net/10037/27041
dc.language.isoengen_US
dc.publisherUiT Norges arktiske universiteten_US
dc.publisherUiT The Arctic University of Norwayen_US
dc.relation.haspart<p>Paper I: Wickstrøm, K., Kampffmeyer, M.C. & Jenssen, R. (2020). Uncertainty and interpretability in convolutional neural networks for semantic segmentation of colorectal polyps. <i>Medical Image Analysis, 60</i>, 101619. Also available in Munin at <a href=https://hdl.handle.net/10037/17135>https://hdl.handle.net/10037/17135</a>. <p>Paper II: Wickstrøm, K.K., Mikalsen, K.Ø., Kampffmeyer, M.C., Revhaug, A. & Jenssen, R. (2021). Uncertainty-aware deep ensembles for reliable and explainable predictions of clinical time series. <i>IEEE Journal of Biomedical and Health Informatics, 25</i>(7), 2435 - 2444. Published version not available in Munin due to publisher’s restrictions. Published version available at <a href=https://doi.org/10.1109/JBHI.2020.3042637>https://doi.org/10.1109/JBHI.2020.3042637</a>. Accepted manuscript version available in Munin at <a href=https://hdl.handle.net/10037/27037>https://hdl.handle.net/10037/27037</a>. <p>Paper III: Wickstrøm, K.K., Trosten, D.J., Løkse, S., Boubekki, A., Mikalsen, K.Ø., Kampffmeyer, M.C. & Jenssen, R. RELAX: representation learning explainability. (Submitted manuscript). Also available in arXiv at <a href=https://doi.org/10.48550/arXiv.2112.10161>https://doi.org/10.48550/arXiv.2112.10161</a>. <p>Paper IV: Wickstrøm, K.K., Østmo, E.A., Radiya, K., Mikalsen, K.Ø., Kampffmeyer, M.C. & Jenssen, R. A clinically motivated self-supervised approach for content-based image retrieval of CT liver images. (Submitted manuscript). Also available in arXiv at <a href= https://doi.org/10.48550/arXiv.2207.04812>https://doi.org/10.48550/arXiv.2207.04812</a>.en_US
dc.rights.accessRightsembargoedAccessen_US
dc.rights.holderCopyright 2022 The Author(s)
dc.subject.courseIDDOKTOR-004
dc.subjectVDP::Mathematics and natural science: 400::Information and communication science: 420::Simulation, visualization, signal processing, image processing: 429en_US
dc.subjectVDP::Matematikk og Naturvitenskap: 400::Informasjons- og kommunikasjonsvitenskap: 420::Simulering, visualisering, signalbehandling, bildeanalyse: 429en_US
dc.subjectVDP::Mathematics and natural science: 400::Information and communication science: 420::Knowledge based systems: 425en_US
dc.subjectVDP::Matematikk og Naturvitenskap: 400::Informasjons- og kommunikasjonsvitenskap: 420::Kunnskapsbaserte systemer: 425en_US
dc.titleAdvancing Deep Learning with Emphasis on Data-Driven Healthcareen_US
dc.typeDoctoral thesisen_US
dc.typeDoktorgradsavhandlingen_US


Tilhørende fil(er)

Thumbnail
Thumbnail
Thumbnail

Denne innførselen finnes i følgende samling(er)

Vis enkel innførsel