Dense Embeddings from Self-Supervision and Foundation Models Improve Cell Linking Performance

Constantin Dalinghaus, Anwai Archit, Constantin Pape

Keywords: cell tracking, microscopy, self-supervision features, foundation models

Registration Requirement: Yes

Abstract: Cell tracking is an important task in microscopy, enabling the study of cell population dynamics. The state-of-the-art uses tracking-by-detection and has substantially improved due to advances in cell segmentation and transformer-based cell linking. How can these systems be further improved once cell segmentation performance plateaus? A promising avenue is to enhance the shallow features used in cell linking with learned features. We investigate this approach by integrating learned features from self-supervised learning and foundation models within Trackastra, showing improved performance on two tracking datasets.

Reproducibility: https://github.com/orgs/MIDL26-Short-Tracking/repositories

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 126