SAC-Net: Learning with weak and noisy labels in histopathology image segmentation

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Deep Learning

Background:

• Deep convolutional neural networks excel at image segmentation but struggle with complex instances like nuclei in histopathology.
• Weakly supervised learning reduces annotation burden but faces a performance gap compared to fully supervised methods.

Purpose of the Study:

• To develop a weakly-supervised nuclei segmentation method requiring only nuclear centroid annotations.
• To improve the accuracy and efficiency of nuclei segmentation in histopathology images.

Main Methods:

• A two-stage training approach using a segmentation network (SAC-Net) with constraint and attention mechanisms.
• Generation of pseudo ground truth labels (boundary and superpixel masks) for initial training.
• Refinement of pseudo labels using Confident Learning for pixel-level accuracy.

Main Results:

• The proposed method demonstrates highly competitive performance in cell nuclei segmentation.
• Validation on three public histopathology image datasets confirms the method's effectiveness.
• The approach successfully addresses challenges posed by noisy labels in weakly supervised learning.

Conclusions:

• The centroid-based weakly supervised nuclei segmentation method offers a viable alternative to fully supervised approaches.
• The two-stage training strategy with Confident Learning refinement enhances segmentation accuracy.
• This work contributes to advancing automated nuclei segmentation in digital pathology.