Weakly supervised semantic segmentation of leukocyte images based on class activation maps

Area of Science:

• Medical Image Analysis
• Computational Biology
• Computer Vision

Background:

• Accurate segmentation of leukocyte images is vital for automated detection within the human defense system.
• Traditional fully supervised semantic segmentation (FSSS) requires extensive, labor-intensive pixel-level annotations.
• Existing methods face challenges with time and cost associated with manual data labeling.

Purpose of the Study:

• To develop a weakly supervised semantic segmentation (WSSS) approach for leukocyte images.
• To reduce the reliance on time-consuming pixel-level annotations.
• To achieve high segmentation accuracy comparable to FSSS methods.

Main Methods:

• Utilized improved Class Activation Maps (CAMs) for WSSS.
• Employed image preprocessing to enhance leukocyte boundaries.
• Integrated an attention mechanism to refine CAMs by matching local and global features.
• Leveraged random walks, dense conditional random fields, and hole filling for pseudo-label generation.
• Trained a UNet model using the generated pseudo-segmentation labels.

Main Results:

• The proposed WSSS method effectively generates pseudo-segmentation labels for leukocyte images.
• Training a UNet with these pseudo-labels achieved segmentation performance close to FSSS.
• Demonstrated significant reduction in manual annotation costs.
• Validated performance on BCCD and TMAMD datasets.

Conclusions:

• The developed WSSS approach offers an efficient alternative to FSSS for leukocyte image segmentation.
• This method successfully reduces annotation burden while maintaining high segmentation accuracy.
• The technique holds promise for advancing automated leukocyte detection systems.