ASIST: Annotation-free synthetic instance segmentation and tracking by adversarial simulations.

Area of Science:

• Microscopy image analysis
• Computational biology
• Computer vision

Background:

• Quantitative analysis of microscope videos requires instance segmentation and tracking of cellular/subcellular objects.
• Traditional methods use two stages: segmentation then frame-by-frame tracking.
• Deep learning offers single-stage solutions but demands extensive, consistent spatial and temporal annotations.

Purpose of the Study:

• To develop an annotation-free method for instance segmentation and tracking in microscopy videos.
• To address the challenges of dense and dynamic objects in microscopy data.
• To reduce the resource-intensive nature of data annotation for deep learning models.

Main Methods:

• Proposed the annotation-free synthetic instance segmentation and tracking (ASIST) method.
• Integrated adversarial simulation with single-stage pixel-embedding based deep learning.
• Evaluated the method on both cellular (HeLa) and subcellular (microvilli) objects.

Main Results:

• ASIST demonstrated 7%-11% higher performance in segmentation, detection, and tracking for microvilli compared to fully supervised methods.
• Achieved comparable performance on HeLa cell videos.

Conclusions:

• The ASIST method represents a significant advancement in annotation-free instance segmentation and tracking for microscopy.
• This is the first study to explore annotation-free approaches for this task in microscopy videos.
• The method effectively combines adversarial simulation and pixel-embedding learning to overcome annotation limitations.