Contour-Seed Pairs Learning-Based Framework for Simultaneously Detecting and Segmenting Various Overlapping...

Area of Science:

• Biomedical Image Analysis
• Computational Pathology
• Cell Biology

Background:

• Automated segmentation of granular objects in microscopy is crucial for clinical applications like cancer grading.
• Existing methods often focus on specific cell types or struggle with inferring contours of clustered objects.

Purpose of the Study:

• To develop a robust and automated framework for cell/nucleus segmentation using contour-seed pairs learning.
• To address limitations in current methods by unifying detection and segmentation into a single regression problem.

Main Methods:

• A novel contour-seed pairs learning framework based on a cascade sparse regression chain model.
• Iterative updating of object locations and boundaries using learned convolutional features.
• Contour inference optimized by the minimum description length principle.

Main Results:

• The proposed method effectively delineates boundaries of individual cells, including overlapping and complex shapes.
• Achieved state-of-the-art performance on four challenging datasets: kidney renal cell carcinoma, Drosophila Kc167, red blood cells, and cervical cytology.
• Demonstrated strong generic performance across various cell/nucleus types.

Conclusions:

• The developed framework offers a generic and robust solution for automated cell/nucleus segmentation and contour inference.
• This approach significantly advances the field of biomedical image analysis for pathological and biological studies.