Improving Data-Efficiency and Robustness of Medical Imaging Segmentation Using Inpainting-Based Self-Supervised Learning.
CT
MRI
deep learning
machine learning
segmentation
self-supervised learning
Journal
Bioengineering (Basel, Switzerland)
ISSN: 2306-5354
Titre abrégé: Bioengineering (Basel)
Pays: Switzerland
ID NLM: 101676056
Informations de publication
Date de publication:
04 Feb 2023
04 Feb 2023
Historique:
received:
21
12
2022
revised:
16
01
2023
accepted:
23
01
2023
entrez:
25
2
2023
pubmed:
26
2
2023
medline:
26
2
2023
Statut:
epublish
Résumé
We systematically evaluate the training methodology and efficacy of two inpainting-based pretext tasks of context prediction and context restoration for medical image segmentation using self-supervised learning (SSL). Multiple versions of self-supervised U-Net models were trained to segment MRI and CT datasets, each using a different combination of design choices and pretext tasks to determine the effect of these design choices on segmentation performance. The optimal design choices were used to train SSL models that were then compared with baseline supervised models for computing clinically-relevant metrics in label-limited scenarios. We observed that SSL pretraining with context restoration using 32 × 32 patches and Poission-disc sampling, transferring only the pretrained encoder weights, and fine-tuning immediately with an initial learning rate of 1 × 10-3 provided the most benefit over supervised learning for MRI and CT tissue segmentation accuracy (
Identifiants
pubmed: 36829701
pii: bioengineering10020207
doi: 10.3390/bioengineering10020207
pmc: PMC9951871
pii:
doi:
Types de publication
Journal Article
Langues
eng
Subventions
Organisme : NIBIB NIH HHS
ID : R01 EB002524
Pays : United States
Organisme : National Science Foundation
ID : DGE 1656518
Organisme : NIH HHS
ID : P41 EB027060
Pays : United States
Organisme : GE Healthcare
ID : NA
Organisme : National Science Foundation
ID : CCF1563078
Organisme : NIH HHS
ID : R01 AR079431
Pays : United States
Organisme : Philips (United States)
ID : NA
Organisme : NIH HHS
ID : 75N92020C00021
Pays : United States
Organisme : United States Department of Defense
ID : National Science and Engineering Graduate Fellowship (ARO)
Organisme : National Science Foundation
ID : CCF1763315
Organisme : Stanford University
ID : Precision Health and Integrated Diagnostics Seed Grant
Organisme : NIH HHS
ID : R01 EB002524
Pays : United States
Organisme : NIAMS NIH HHS
ID : R01 AR077604
Pays : United States
Organisme : NIAMS NIH HHS
ID : K24 AR062068
Pays : United States
Organisme : NIBIB NIH HHS
ID : P41 EB027060
Pays : United States
Organisme : NIH HHS
ID : R01 AR077604
Pays : United States
Organisme : NIH HHS
ID : 75N92020C00008
Pays : United States
Organisme : NIH HHS
ID : K24 AR062068
Pays : United States
Références
AJR Am J Roentgenol. 2021 Jun;216(6):1614-1625
pubmed: 32755384
IEEE J Biomed Health Inform. 2022 Nov;26(11):5310-5320
pubmed: 34478389
J Magn Reson Imaging. 2019 Jun;49(7):e183-e194
pubmed: 30582251
IEEE Trans Pattern Anal Mach Intell. 2021 Nov;43(11):4037-4058
pubmed: 32386141
Magn Reson Imaging. 2017 May;38:63-70
pubmed: 28017730
J Magn Reson Imaging. 2022 Jun;55(6):1650-1663
pubmed: 34918423
Eur Radiol. 2020 Apr;30(4):2231-2240
pubmed: 31844957
J Magn Reson Imaging. 2020 Nov;52(5):1321-1339
pubmed: 31755191
IEEE Trans Med Imaging. 2021 Dec;40(12):3543-3554
pubmed: 34138702
Med Image Anal. 2019 Dec;58:101539
pubmed: 31374449
Radiol Artif Intell. 2021 Feb 10;3(3):e200078
pubmed: 34235438
J Magn Reson Imaging. 2018 May;47(5):1328-1341
pubmed: 29090500
Radiol Clin North Am. 2022 Jul;60(4):575-582
pubmed: 35672090
Med Image Anal. 2021 Apr;69:101950
pubmed: 33421920
Nat Methods. 2020 Mar;17(3):261-272
pubmed: 32015543
IEEE Trans Med Imaging. 2020 Aug;39(8):2626-2637
pubmed: 32730213
JAMA Netw Open. 2019 Jun 5;2(6):e195600
pubmed: 31173130