Learning from crowds for automated histopathological image segmentation.

Area of Science:

• Computational Pathology
• Medical Image Analysis
• Machine Learning

Background:

• Automated semantic segmentation of histopathological images is crucial for Computational Pathology (CPATH).
• Deep Learning (DL) methods are limited by the scarcity of expert annotations.
• Crowdsourcing (CR) offers a solution but faces challenges with noisy, non-expert annotations.

Purpose of the Study:

• To develop a scalable crowdsourcing approach for histopathological image segmentation.
• To jointly learn expert segmentation and annotator expertise without training a separate model per annotator.
• To address the limitations of existing methods that do not scale with a large number of annotators.

Main Methods:

• Proposed a novel family of methods using two coupled neural networks: a segmentation network and an annotator network.
• The annotator network estimates annotator behavior using a single network that takes annotator ID as input, enabling scalability.
• Introduced a new annotator network model considering global image features for improved expertise estimation.

Main Results:

• The proposed CR modeling achieved a Dice coefficient of 0.7827 on Triple Negative Breast Cancer images.
• Outperformed the established STAPLE algorithm (0.7039).
• Demonstrated competitive performance compared to supervised methods using expert labels (0.7723).

Conclusions:

• The developed coupled network approach offers a scalable and effective solution for crowdsourced histopathological image segmentation.
• This method successfully models annotator expertise and segmentation, overcoming limitations of previous non-scalable approaches.
• The findings suggest a viable path for leveraging crowdsourcing in real-world Computational Pathology applications.