S2L-CM: Scribble-supervised nuclei segmentation in histopathology images using contrastive regularization and pixel-level multiple instance learning

Deep learning-based pathology nuclei segmentation algorithms have demonstrated remarkable performance. Conventional methods mostly focus on supervised learning, which requires significant manual effort to generate ground truth labels. Recently, weakly supervised learning has been extensively explore...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Computers in biology and medicine Jg. 192; H. Pt B; S. 110293
Hauptverfasser: Oh, Hyun-Jic, Min, Seonghui, Jeong, Won-Ki
Format: Journal Article
Sprache:Englisch
Veröffentlicht: United States Elsevier Ltd 01.06.2025
Schlagworte:
Algorithms
Cell Nucleus - pathology
Deep Learning
Humans
Image Processing, Computer-Assisted - methods
Internal Medicine
Multiple-Instance Learning Algorithms
Multiscale contrastive regularization
Other
Pixel-level multiple instance learning
Pseudo label supervision
Weakly supervised nuclei segmentation
Pixel-level multiple instance learning
Weakly supervised nuclei segmentation
Pseudo label supervision
Multiscale contrastive regularization
ISSN:0010-4825, 1879-0534, 1879-0534
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Deep learning-based pathology nuclei segmentation algorithms have demonstrated remarkable performance. Conventional methods mostly focus on supervised learning, which requires significant manual effort to generate ground truth labels. Recently, weakly supervised learning has been extensively explored as a method for overcoming this limitation by training models with sparse annotations. However, the performance is inferior compared to that of fully supervised learning schemes. This paper proposes S2L-CM, a scribble-supervised nuclei segmentation framework based on two ideas; first, we leverage self-generated pseudo labels from user-given sparse scribble labels to train the deep learning model without full ground-truth labels, and second, we utilize multiscale contrastive regularization and pixel-level multiple-instance learning to further refine pseudo labels to improve segmentation performance. We demonstrate the effectiveness and robustness of our method on four nuclei datasets by comparing it with existing state-of-the-art methods. Code will be available at: https://github.com/hvcl/S2L-CM. [Display omitted] •Nuclei segmentation is vital for pathology image analysis but relies on labeled data.•Weakly supervised learning reduces data dependency while ensuring good performance.•Using model predictions as pseudo labels enhances model performance.•Multiscale contrastive regularization controls feature distances using prediction.•Pixel-level multiple instance learning contributes to reduce false predictions.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 23
ISSN:0010-4825
1879-0534
1879-0534
DOI:10.1016/j.compbiomed.2025.110293