Learn to segment single cells with deep distance estimator and deep cell detector

Area of Science:

• Cell Biology
• Computational Biology
• Image Analysis

Background:

• Single cell segmentation is crucial for cell imaging analysis but challenging due to data limitations and image characteristics.
• Traditional methods are labor-intensive and lack transferability.
• Deep convolutional neural networks (CNNs) offer efficient segmentation but struggle with specific cell imaging data issues.

Purpose of the Study:

• To develop an improved single cell segmentation strategy overcoming limitations of existing methods.
• To enhance cell counting accuracy, particularly for challenging image conditions.

Main Methods:

• A hybrid approach combining CNNs with the watershed algorithm was developed.
• A CNN (deep distance estimator) was trained to predict the Euclidean distance transform (EDT) of cell masks.
• A faster R-CNN (deep cell detector) was trained to identify cells in the EDT image, followed by watershed segmentation.

Main Results:

• The combined method achieved comparable pixel-wise accuracy to pure CNN approaches.
• Significantly higher cell count accuracy was observed with the combined method, especially for connected and blurry cells.
• Both deep learning components (deep distance estimator and deep cell detector) demonstrated fast convergence and ease of training.

Conclusions:

• The hybrid CNN-watershed approach offers superior cell counting accuracy over traditional and pure CNN pixel-wise methods.
• This strategy effectively addresses challenges posed by noisy, densely packed cell images.
• The developed method is robust, efficient, and easier to train for cell segmentation tasks.