DMNet

DMNet performs cell segmentation and multi-object tracking in microscopy video sequences to resolve dense, touching cells and indistinct boundaries using marker guidance and association-based tracking.

Key Features:

• Dual-Stream Architecture: An explicit cell marker-detection stream guides a mask-prediction stream to localize cells and inform segmentation decisions.
• Distance Map Penalty and Supervised Training: A distance map penalty function is incorporated during supervised training to emphasize separation of touching or closely situated cells.
• M2Track Tracking-by-Detection: Multi-object tracking uses the M2Track approach with short-term track-to-cell associations and track-to-track associations to re-link tracklets with missing segmentation masks across frames.

Scientific Applications:

• Clinical Diagnostics: Accurate cell segmentation and tracking in microscopy videos to support diagnosis and study of disease-related cellular behavior.
• Biomedical Research: Analysis of dense cell populations and cells with indistinct boundaries in time-lapse microscopy for studies of cell dynamics.
• Benchmarking: Demonstrated multiple top-three rankings in the IEEE ISBI 2021 6th Cell Tracking Challenge (CTC-6) across diverse cell types.

Methodology:

DMNet combines an explicit marker-detection stream and a mask-prediction stream with a distance map penalty during supervised training, and applies the M2Track tracking-by-detection pipeline using short-term track-to-cell and subsequent track-to-track associations to relink tracklets with missing segmentation masks.