Cell Segmentation Using a Similarity Interface With a Multi-Task Convolutional Neural Network.

Area of Science:

• Computational Biology
• Image Analysis
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are effective for cell segmentation but demand extensive pixel-level annotations.
• Existing methods struggle with overlapping cells and variations in microscopy image contrast and texture.

Purpose of the Study:

• To develop a multitask learning algorithm for simultaneous cell detection and segmentation using CNNs.
• To reduce the need for pixel-level annotations by utilizing dot annotations for cell centroids.
• To enhance adaptability to diverse microscopy image conditions through a novel similarity interface.

Main Methods:

• A multitask CNN architecture sharing convolutional layers for both detection (centroid prediction) and segmentation (foreground/background mapping).
• Utilized dot annotations (approximate cell centroids) for training data generation, significantly reducing annotation effort.
• Introduced a Similarity Interface (SI) with an unsupervised first layer and a Neighborhood Similarity Layer (NSL) for domain adaptation and handling image variability.

Main Results:

• Achieved comparable or superior detection and segmentation scores against state-of-the-art methods.
• Demonstrated improved separation of overlapping cells due to simultaneous multitask learning.
• Significantly reduced the effort required for generating training data compared to traditional methods.

Conclusions:

• The proposed multitask learning approach offers an efficient and effective solution for cell detection and segmentation in microscopy.
• The Similarity Interface enhances model robustness and adaptability across different imaging conditions.
• This method lowers the barrier to entry for cell image analysis by minimizing annotation requirements.