Tracking object instances, such as individual living cells or molecular accumulations, and their behavior is a common challenge in 2D and 3D volumetric biomedical imaging. Combined with dynamic environments caused by unintentional camera misalignment over longitudinal studies, deformable tissues, and morphologically changing instances, object tracking poses a challenging and time-consuming task in biomedical settings. This paper presents a robust 2D and 3D object tracker designed to handle such scenarios, ensuring consistent identification of physical objects over time. Our method combines image registration with a novel graph-based conflict resolution algorithm for object and lineage tracking, handling camera misalignment and object movement. The algorithm accommodates morphological changes including splitting, merging, growing, shrinking, emerging, and vanishing object instances. This training-free framework serves as a post-processing step after instance segmentation, making it compatible with many tracking-by-detection approaches and generalist segmentation models such as CellPose-SAM. The proposed framework provides a flexible, training-free, and interpretable solution for long-term tracking in neuroscience, cell biology, and medical imaging. Our method is evaluated in two biomedical contexts: tracking 3D Beta-amyloid accumulations in in-vivo two-photon fluorescence imaging and monitoring cell movement and proliferation in 2D cultures. On 2D cell-tracking datasets, MOLT achieves TAR scores of (92.618–96.145%) on CellPose-SAM annotations, outperforming established TrackMate baselines on two of three datasets. For 3D Beta-amyloid plaque tracking, MOLT achieves a Lineage Reconstruction score (LNR) of 83.394%, demonstrating accurate lineage reconstruction under complex splitting and merging dynamics.
Bausch et al. (2026) studied this question.