CMTT-JTracker

CMTT-JTracker combines pre-segmentation and multi-task learning with spatial attention to achieve automated cell tracking with balanced computational efficiency and accuracy across diverse cell datasets.

Key Features:

• CMTT Pre-segmentation Mechanism: Extracts target information at different resolutions without additional training.
• Multi-task Learning Network with Spatial Attention: Simultaneously performs detection and re-identification tasks based on CMTT-extracted features.
• Fully Test-time Adaptation: Adapts at test time without requiring retraining for new datasets.
• Segmentation Performance: Achieves Dice coefficients ranging from 0.758 to 0.928 across different datasets, outperforming SOTA segmentation benchmarks particularly in dense cell scenarios.
• Tracking Accuracy: Achieves MOTA of 0.894 on Fluo-N2DH-SIM+ and MOTA of 0.850 on PhC-C2DL-PSC.

Scientific Applications:

• Automated Cell Lineage Construction: Tracks cells to reconstruct lineage relationships in time-lapse microscopy.
• Cellular Activity Monitoring: Enables automated monitoring of cellular behavior and dynamics.
• Dense Cell Population Analysis: Excels in tracking and segmentation scenarios with densely packed cells.

Methodology:

CMTT-JTracker employs a central-metric fully test-time adaptive (CMTT) mechanism for pre-segmentation of cell images, followed by a multi-task learning network with spatial attention that performs simultaneous detection and re-identification.