PS-Net: human perception-guided segmentation network for EM cell membrane.

Area of Science:

• Computer Vision
• Biomedical Imaging
• Neuroscience

Background:

• Cell membrane segmentation in electron microscopy (EM) is vital for image processing.
• Current methods excel on low-resolution EM data but struggle with high-resolution datasets.
• Human vision performs consistently across resolutions, offering insights for improved algorithms.

Purpose of the Study:

• To understand human visual perception in EM image analysis.
• To develop a computational framework for membrane segmentation inspired by human vision.
• To introduce a novel evaluation metric and a deep learning model for enhanced segmentation.

Main Methods:

• Conducted eye movement and perceptual consistency experiments to analyze human visual strategies.
• Developed the perceptual Hausdorff distance (PHD) as a novel evaluation metric.
• Proposed the PHD-guided segmentation network (PS-Net) with a multiscale architecture and adaptive PHD loss functions.

Main Results:

• Human observers prioritize membrane structure over perfect alignment, unlike conventional metrics.
• The PHD metric aligns better with human perception than existing criteria.
• PS-Net demonstrated superior performance over state-of-the-art methods on diverse EM and natural image datasets.

Conclusions:

• Human visual perception offers a valuable model for improving EM image segmentation algorithms.
• The proposed PHD metric and PS-Net framework significantly advance membrane segmentation accuracy and robustness.
• The findings pave the way for more effective analysis of high-resolution EM data.