Cell Instance Segmentation: The Devil Is in the Boundaries.

Area of Science:

• Computational Biology
• Image Analysis
• Deep Learning

Background:

• State-of-the-art cell instance segmentation relies on deep learning semantic segmentation.
• Current methods often lose geometric cell properties by using pixel-wise objectives.
• Existing approaches struggle to accurately distinguish individual cell instances.

Purpose of the Study:

• To introduce a novel pixel clustering method, Ceb (Cell boundaries), for improved cell instance segmentation.
• To leverage cell boundary features and labels for more accurate instance division.
• To address the limitations of pixel-wise objectives in preserving cell geometry.

Main Methods:

• Ceb utilizes probability maps from semantic segmentation and a revised Watershed algorithm to identify boundary candidates.
• A boundary signature is created for each candidate, incorporating foreground-foreground and background-foreground boundary pixels.
• A lightweight boundary classifier predicts binary labels for boundary candidates, enabling region division and merging for instance segmentation.

Main Results:

• Ceb demonstrates superior performance compared to existing pixel clustering methods on semantic segmentation probability maps.
• The method effectively preserves geometric properties like shape, curvature, and convexity of cell instances.
• Ceb achieves highly competitive results against state-of-the-art cell instance segmentation techniques.

Conclusions:

• Ceb offers a novel and effective approach to cell instance segmentation by focusing on boundary information.
• The method overcomes limitations of pixel-wise objectives, enhancing the preservation of crucial cell geometric features.
• Ceb provides a robust and competitive alternative for accurate cell instance segmentation in biological imaging.