Joint fully convolutional and graph convolutional networks for weakly-supervised segmentation of pathology images

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Accurate tissue/region segmentation in pathology images is crucial for quantitative analysis.
• Traditional methods demand pixel-level annotations, which are labor-intensive and difficult to obtain.
• Weakly-supervised approaches offer a promising alternative to overcome annotation limitations.

Purpose of the Study:

• To propose a novel weakly-supervised model, FGNet, for automated segmentation of pathology images.
• To reduce the reliance on pixel-level annotations by utilizing image-level labels.
• To achieve competitive segmentation performance comparable to fully-supervised methods.

Main Methods:

• Developed a joint Fully Convolutional and Graph Convolutional Network (FGNet) model.
• Employed image-level foreground proportion as weak supervision for training.
• Integrated dynamic superpixel operations and an uncertainty range constraint for robust segmentation.

Main Results:

• FGNet demonstrated competitive segmentation results on HER2, KI67, and H&E pathology image datasets.
• The model effectively reduces the need for extensive pixel-level annotations.
• Achieved robust performance through mutable superpixel usage and uncertainty constraints.

Conclusions:

• FGNet offers an effective weakly-supervised solution for pathology image segmentation.
• The proposed method significantly lowers the annotation burden in digital pathology.
• FGNet shows promise for advancing quantitative analysis in cancer research.