Semi-Supervised Histopathology Image Segmentation with Feature Diversified Collaborative Learning

Thanh-Huy Nguyen, Nguyen Lan Vi Vu, Hoang-Thien Nguyen, Quang-Vinh Dinh, Xingjian Li, Min Xu
Proceedings of The First AAAI Bridge Program on AI for Medicine and Healthcare, PMLR 281:165-172, 2025.

Abstract

Histopathology image segmentation plays a critical role in advancing disease diagnosis, prognosis, and treatment planning. However, it presents significant challenges due to the complexity of tissue structures, staining variability, and low contrast between tissue classes. Semi-supervised learning, employed to mitigate annotation scarcity, introduces additional difficulties, such as managing noisy pseudo-labels while ensuring robust performance with limited supervision. Traditional collaborative training methods, commonly used in medical image segmentation, often face issues like model coupling—where models become overly dependent on each other, propagating similar errors—or confirmation bias, where networks reinforce initial mistakes by relying on inaccurate pseudo-labels. Existing frameworks designed to tackle these challenges often suffer from complex pipelines and require extensive pre-training but fail to address the noise characteristics inherent in such datasets. To balance the efficiency of traditional co-training methods with dual networks while enhancing segmentation accuracy on noisy histopathological data, we propose Feature Diversified Collaborative Learning (FDCL). Our work aims to design an effective feature diversification loss that encourages the feature representations of sub-networks to be distinct, ensuring they capture different information to exchange with each other, thereby avoiding suboptimal solutions or, even worse, falling into the coupling problem. We benchmark our method on two well-known histopathology datasets and achieve state-of-the-art results on the GlaS dataset with only 10% of the labeled data. Code is available at https://github.com/vnlvi2k3/FDCL.

Cite this Paper

BibTeX
@InProceedings{pmlr-v281-nguyen25b, title = {Semi-Supervised Histopathology Image Segmentation with Feature Diversified Collaborative Learning}, author = {Nguyen, Thanh-Huy and Vu, Nguyen Lan Vi and Nguyen, Hoang-Thien and Dinh, Quang-Vinh and Li, Xingjian and Xu, Min}, booktitle = {Proceedings of The First AAAI Bridge Program on AI for Medicine and Healthcare}, pages = {165--172}, year = {2025}, editor = {Wu, Junde and Zhu, Jiayuan and Xu, Min and Jin, Yueming}, volume = {281}, series = {Proceedings of Machine Learning Research}, month = {25 Feb}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v281/main/assets/nguyen25b/nguyen25b.pdf}, url = {https://proceedings.mlr.press/v281/nguyen25b.html}, abstract = {Histopathology image segmentation plays a critical role in advancing disease diagnosis, prognosis, and treatment planning. However, it presents significant challenges due to the complexity of tissue structures, staining variability, and low contrast between tissue classes. Semi-supervised learning, employed to mitigate annotation scarcity, introduces additional difficulties, such as managing noisy pseudo-labels while ensuring robust performance with limited supervision. Traditional collaborative training methods, commonly used in medical image segmentation, often face issues like model coupling—where models become overly dependent on each other, propagating similar errors—or confirmation bias, where networks reinforce initial mistakes by relying on inaccurate pseudo-labels. Existing frameworks designed to tackle these challenges often suffer from complex pipelines and require extensive pre-training but fail to address the noise characteristics inherent in such datasets. To balance the efficiency of traditional co-training methods with dual networks while enhancing segmentation accuracy on noisy histopathological data, we propose Feature Diversified Collaborative Learning (FDCL). Our work aims to design an effective feature diversification loss that encourages the feature representations of sub-networks to be distinct, ensuring they capture different information to exchange with each other, thereby avoiding suboptimal solutions or, even worse, falling into the coupling problem. We benchmark our method on two well-known histopathology datasets and achieve state-of-the-art results on the GlaS dataset with only 10% of the labeled data. Code is available at https://github.com/vnlvi2k3/FDCL.} }
Endnote
%0 Conference Paper %T Semi-Supervised Histopathology Image Segmentation with Feature Diversified Collaborative Learning %A Thanh-Huy Nguyen %A Nguyen Lan Vi Vu %A Hoang-Thien Nguyen %A Quang-Vinh Dinh %A Xingjian Li %A Min Xu %B Proceedings of The First AAAI Bridge Program on AI for Medicine and Healthcare %C Proceedings of Machine Learning Research %D 2025 %E Junde Wu %E Jiayuan Zhu %E Min Xu %E Yueming Jin %F pmlr-v281-nguyen25b %I PMLR %P 165--172 %U https://proceedings.mlr.press/v281/nguyen25b.html %V 281 %X Histopathology image segmentation plays a critical role in advancing disease diagnosis, prognosis, and treatment planning. However, it presents significant challenges due to the complexity of tissue structures, staining variability, and low contrast between tissue classes. Semi-supervised learning, employed to mitigate annotation scarcity, introduces additional difficulties, such as managing noisy pseudo-labels while ensuring robust performance with limited supervision. Traditional collaborative training methods, commonly used in medical image segmentation, often face issues like model coupling—where models become overly dependent on each other, propagating similar errors—or confirmation bias, where networks reinforce initial mistakes by relying on inaccurate pseudo-labels. Existing frameworks designed to tackle these challenges often suffer from complex pipelines and require extensive pre-training but fail to address the noise characteristics inherent in such datasets. To balance the efficiency of traditional co-training methods with dual networks while enhancing segmentation accuracy on noisy histopathological data, we propose Feature Diversified Collaborative Learning (FDCL). Our work aims to design an effective feature diversification loss that encourages the feature representations of sub-networks to be distinct, ensuring they capture different information to exchange with each other, thereby avoiding suboptimal solutions or, even worse, falling into the coupling problem. We benchmark our method on two well-known histopathology datasets and achieve state-of-the-art results on the GlaS dataset with only 10% of the labeled data. Code is available at https://github.com/vnlvi2k3/FDCL.
APA
Nguyen, T., Vu, N.L.V., Nguyen, H., Dinh, Q., Li, X. & Xu, M.. (2025). Semi-Supervised Histopathology Image Segmentation with Feature Diversified Collaborative Learning. Proceedings of The First AAAI Bridge Program on AI for Medicine and Healthcare, in Proceedings of Machine Learning Research 281:165-172 Available from https://proceedings.mlr.press/v281/nguyen25b.html.

Related Material