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
On Medical Image Segmentation With Noisy Labels
KTH, School of Engineering Sciences (SCI), Mathematics (Dept.), Mathematical Statistics.ORCID iD: 0009-0006-4292-828X
2023 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

It is well known that data sets used for training and testing automatic medical image segmentation methods often contain a lot of label noise. Such noise affects the performance of the methods and has been subject to a lot of research. One way to approach the topic of label noise that largely has been overlooked in the literature is to investigate how it affects the theoretically optimal segmentations. This thesis consist of four papers related to such investigations for the two most popular choices of loss functions in the field, cross-entropy and soft-Dice, and the most popular metric, Dice.

In paper A, the loss functions cross-entropy and soft-Dice are investigated. Inspired by work from binary classification, the properties of calibration and convexity are proposed to explain the experimental observations of good stability associated with cross-entropy and good performance associated with soft-Dice. It is then shown that soft-Dice neither is convex nor quasi-convex and it is conjectured that soft-Dice is calibrated to Dice. Finally, an alternative quasi-convex loss function is experimentally compared to soft-Dice on a kidney segmentation problem.

In paper B, the optimal segmentations to the metrics Accuracy and Dice are characterized when noise is present. This characterization is then used to give a detailed account of the volume bias associated with the metrics. In particular, sharp bounds for volume bias is provided, it is shown that the volume of an optimizer to Accuracy always is less than or equal to the volume of an optimizer to Dice and that the set of optimizers to the two metrics coincide when the optimization is constrained to the segmentations without volume bias. Finally, experimental results supporting the observations are presented on a set of segmentation problems.

In paper C, the effect label noise has on soft-Dice is studied. In particular, the optimal solutions are characterized and sharp bounds for the volume biasis provided. Moreover the conjecture of soft-Dice being calibrated to Dice is proved under a compactness assumption that always holds in practice. Finally, experimental results supporting the observations are presented on a set of organ segmentation problems and a set of synthetic segmentation problems.

In paper D, a noise model based on Gaussian field deformations is proposed. Several theoretical properties for labels with this sort of noise is proved, including a closed form expression for marginal probabilities and a representation that can be used for efficient sampling. The noise model is then used to study 1/2-thresholded optimal solutions to the loss functions cross-entropyand soft-Dice, and it is shown how they diverge as the noise level is increased. Finally, by using the characterization of the optimal solutions to soft-Dice it is shown how cross-entropy can be used in conjunction with an a priori unknown but computable threshold to recover optimal solutions to soft-Dice. The theoretical observations are validated on three organ segmentation problems with various levels of noise. 

Abstract [sv]

Det är väl känt att data som används för träning och testning av medicinska bildsegmenteringsmetoder ofta innehåller mycket annoteringsbrus. Sådant brus påverkar prestandan av metoderna och har givit upphov till en stor mängd forskning. Ett sätt att angripa frågan om annoteringsbrus som i stort har förbisetts är att undersöka hur det påverkar vilka segmenteringar som är teoretiskt optimala. Den här avhandlingen består av fyra artiklar relaterade till sådana studier för de två inom fältet mest populära målfunktionerna cross-entropy och soft-Dice, samt den mest populära metriken Dice.

I artikel A studeras målfunktionen cross-entropy och soft-Dice. Det föreslås, inspirerat av arbete inom binärklassifikation, att konvexitet och kalibrering kan användas för att beskriva de experimentella observationerna av bra stabilitet kopplad till cross-entropy och bra prestanda kopplad till soft-Dice. Sedan bevisas det att soft-Dice vare sig är konvex eller kvasi-konvex och så förmodas det att soft-Dice är kalibrerad till Dice. Slutligen introduceras en alternativ kvasi-konvex målfunktion som experimentellt jämförs med soft-Dice på ett njursegmenteringsproblem.

I artikel B karakteriseras de optimala lösningarna till Accuracy och Dice när brus är närvarande. Denna karakteriseringen används sedan för att ge en detaljerad beskrivning av volymbias associerad med metrikerna. Specifikt visas skarpa gränser för volymbias, att Accuracy-optimala segmenteringar alltid har mindre eller lika stor volym som Dice-optimala segmenteringar samt att de optimala lösningarna till metrikerna sammanfaller när optimeringen görs över mängden av segmenteringar utan volymbias. Slutligen presenteras experimentella resultat som bekräftar de teoretiska observationerna på en uppsättning segmenteringsproblem.

I artikel C undersöks effekten brus har på soft-Dice. De optimala lösningarna karakteriseras och skarpa gränser för volymbias presenteras. Vidare bevisas den tidigare förmodan om att soft-Dice är kalibrerad till Dice under ett kompakthetsantagande som alltid håller i praktiken. Slutligen presenteras experimentella resultat som bekräftar de teoretiska observationerna på en uppsättning organsegmenteringsproblem och en uppsättning syntetiska segmenteringsproblem.

I artikel D presenteras en brusmodell baserad på deformationer av Gaussiska fält. Flera teoretiska egenskaper för brusmodellen bevisas, inklusive en sluten form för marginalsannolikheter samt en representation som kan användas för effektiv simulering. Brusmodellen används sedan för att studera 1/2-trösklade optimala lösningar till cross-entropy och soft-Dice, och specifikt, hur dessa divergerar när nivån av brus går upp. Med hjälp av karakteriseringen av de optimala lösningarna till soft-Dice visas det hur optimala lösningar till soft-Dice kan fås med hjälp av cross-entropy och ett annat tröskelvärde som inte är a priori känt men som effektivt kan beräknas. Till sist valideras de teoretiska observationerna på tre organsegmenteringsproblem.

Place, publisher, year, edition, pages
Stockholm, Sweden: KTH Royal Institute of Technology, 2023. , p. 162
Series
TRITA-SCI-FOU ; 2023;11
Keywords [en]
Medical image segmentation, image segmentation, machine learning, supervised learning, label noise, Sörensen-Dice coefficient, soft-Dice
Keywords [sv]
Medicinsk bildsegmentering, bildsegmentering, maskininlärning, övervakat lärande, annoteringsbrus, Sörensen-Dice koefficienten, soft-Dice
National Category
Medical Imaging
Research subject
Applied and Computational Mathematics; Applied and Computational Mathematics, Mathematical Statistics
Identifiers
URN: urn:nbn:se:kth:diva-326552ISBN: 978-91-8040-538-6 (print)OAI: oai:DiVA.org:kth-326552DiVA, id: diva2:1754892
Public defence
2023-06-01, Kollegiesalen, Brinellvägen 6, Stockholm, 13:00 (Swedish)
Opponent
Supervisors
Note

QC 2023-05-10

Available from: 2023-05-10 Created: 2023-05-04 Last updated: 2025-02-09Bibliographically approved
List of papers
1. Calibrated Surrogate Maximization of Dice
Open this publication in new window or tab >>Calibrated Surrogate Maximization of Dice
Show others...
2020 (English)In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part IV, Springer Nature , 2020, Vol. 12264, p. 269-278Conference paper, Published paper (Refereed)
Abstract [en]

In the medical imaging community, it is increasingly popular to train machine learning models for segmentation problems with objectives based on the soft-Dice surrogate. While experimental studies have showed good performance with respect to Dice, there have also been reports of some issues related to stability. In parallel with these developments, direct optimization of evaluation metrics has also been studied in the context of binary classification. Recently, in this setting, a quasi-concave, lower-bounded and calibrated surrogate for the F1-score has been proposed. In this work, we show how to use this surrogate in the context of segmentation. We then show that it has some better theoretical properties than soft-Dice. Finally, we experimentally compare the new surrogate with soft-Dice on a 3D-segmentation problem and get results indicating that stability is improved. We conclude that the new surrogate, for theoretical and experimental reasons, can be considered a promising alternative to the soft-Dice surrogate.

Place, publisher, year, edition, pages
Springer Nature, 2020
Series
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), ISSN 0302-9743 ; 12264
Keywords
Calibration, Dice, Segmentation
National Category
Medical Imaging
Identifiers
urn:nbn:se:kth:diva-291720 (URN)10.1007/978-3-030-59719-1_27 (DOI)2-s2.0-85092758157 (Scopus ID)
Conference
23rd International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2020; Lima; Peru; 4 October 2020 through 8 October 2020
Note

QC 20210325

Available from: 2021-03-25 Created: 2021-03-25 Last updated: 2025-02-09Bibliographically approved
2. On image segmentation with noisy labels: characterization and volume properties of the optimal solutions to accuracy and dice.
Open this publication in new window or tab >>On image segmentation with noisy labels: characterization and volume properties of the optimal solutions to accuracy and dice.
2022 (English)Conference paper, Published paper (Refereed)
Abstract [en]

We study two of the most popular performance metrics in medical image segmentation, Accuracy and Dice, when the target labels are noisy. For both metrics,several statements related to characterization and volume properties of the set ofoptimal segmentations are proved, and associated experiments are provided. Ourmain insights are: (i) the volume of the solutions to both metrics may deviatesignificantly from the expected volume of the target, (ii) the volume of a solutionto Accuracy is always less than or equal to the volume of a solution to Dice and(iii) the optimal solutions to both of these metrics coincide when the set of feasiblesegmentations is constrained to the set of segmentations with the volume equal tothe expected volume of the target.

National Category
Probability Theory and Statistics
Research subject
Applied and Computational Mathematics, Mathematical Statistics; Computer Science
Identifiers
urn:nbn:se:kth:diva-325140 (URN)
Conference
NeuRIPS 2022
Note

QC 20230403

Available from: 2023-03-31 Created: 2023-03-31 Last updated: 2023-05-04Bibliographically approved
3. Noisy Image Segmentation With Soft-Dice
Open this publication in new window or tab >>Noisy Image Segmentation With Soft-Dice
(English)Manuscript (preprint) (Other academic)
Abstract [en]

This paper presents a study on the soft-Dice loss, one ofthe most popular loss functions in medical image segmentation, for situations where noise is present in target labels. In particular, the set of optimal solutions are characterizedand sharp bounds on the volume bias of these solutions areprovided. It is further shown that a sequence of soft segmentations converging to optimal soft-Dice also converges tooptimal Dice when converted to hard segmentations usingthresholding. This is an important result because soft-Diceis often used as a proxy for maximizing the Dice metric.Finally, experiments confirming the theoretical results areprovided.

National Category
Medical Imaging
Identifiers
urn:nbn:se:kth:diva-326549 (URN)
Note

QC 20230508

Available from: 2023-05-04 Created: 2023-05-04 Last updated: 2025-02-09Bibliographically approved
4. Marginal Thresholding in Noisy Image Segmentation
Open this publication in new window or tab >>Marginal Thresholding in Noisy Image Segmentation
(English)Manuscript (preprint) (Other academic)
Abstract [en]

This work presents a study on label noise in medical image segmentation by considering a noise model based onGaussian field deformations. Such noise is of interest because it yields realistic looking segmentations and because itis unbiased in the sense that the expected deformation is theidentity mapping. Efficient methods for sampling and closedform solutions for the marginal probabilities are provided.Moreover, theoretically optimal solutions to the loss functions cross-entropy and soft-Dice are studied and it is shownhow they diverge as the level of noise increases. Based on recent work on loss function characterization, it is shown thatoptimal solutions to soft-Dice can be recovered by thresholding solutions to cross-entropy with a particular a prioriunknown threshold that efficiently can be computed. Thisraises the question whether the decrease in performanceseen when using cross-entropy as compared to soft-Dice iscaused by using the wrong threshold. The hypothesis is validated in 5-fold studies on three organ segmentation problemsfrom the TotalSegmentor data set, using 4 different strengthsof noise. The results show that changing the threshold leadsthe performance of cross entropy to go from systematicallyworse than soft-Dice to similar or better results than soft-Dice. 

National Category
Medical Imaging
Identifiers
urn:nbn:se:kth:diva-326550 (URN)
Note

QC 20230508

Available from: 2023-05-04 Created: 2023-05-04 Last updated: 2025-02-09Bibliographically approved

Open Access in DiVA

fulltext(4478 kB)630 downloads
File information
File name FULLTEXT01.pdfFile size 4478 kBChecksum SHA-512
d26427d9a51f24cdc34e0cdca9e315adbb6f301d2da7f965ae9dcfa0e072e778644b51516a8d469d2b113f3f7852a597ecc8df0e7b730b0b39b70291aa66f95c
Type fulltextMimetype application/pdf

Search in DiVA

By author/editor
Nordström, Marcus
By organisation
Mathematical Statistics
Medical Imaging

Search outside of DiVA

GoogleGoogle Scholar
Total: 631 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

isbn
urn-nbn

Altmetric score

isbn
urn-nbn
Total: 2140 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