Detecting overlapping instances in microscopy images using extremal region trees.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Machine Learning

Background:

• Microscopy images often present challenges with varying cell densities, impacting automated analysis.
• Existing methods struggle to accurately detect and count cells in both low and high-density regions simultaneously.

Purpose of the Study:

• To develop a unified automated method for detecting and counting all instances of objects in microscopy images, regardless of density or overlap.
• To address the limitations of current cell detection and counting approaches in handling diverse microscopy data.

Main Methods:

• Introduced a tree-structured discrete graphical model for selecting and labeling non-overlapping image regions.
• Utilized structured output Support Vector Machines (SVM) for learning and dynamic programming for inference on a tree-structured region graph.
• Employed weak annotations (a single dot per instance) for training and identified candidate regions using extremal regions of a computed surface.

Main Results:

• The proposed model successfully detects and counts objects in images with partially overlapping and clustered instances.
• Demonstrated improved performance by using a proxy problem for learning the surface to enhance extremal region selection.
• Achieved state-of-the-art performance across six diverse microscopy datasets, including fluorescence, phase contrast, and histopathology images.

Conclusions:

• The developed tree-structured graphical model offers a robust solution for automated cell detection and counting in challenging microscopy images.
• The method's ability to handle varying cell densities and weak annotations makes it a versatile tool for biological research.
• This approach significantly advances the state-of-the-art in automated image analysis for microscopy applications.