Instance segmentation of overlapping cells in microscopy remains a tough challenge due to semi-transparent structures that blur boundaries and create mixed visual evidence. Traditional methods rely on local regions or shape priors, but they often lack global reasoning across overlapping objects.
To address this, researchers present QCell, a novel query-based model that de-overlaps cell instances. The model introduces two key components: an instance recombination module that decomposes and recombines query representations in latent space, and a contrastive query alignment objective that enhances feature learning and separates overlapping queries.
The team also releases a new Organoid dataset for overlapping cell segmentation in brightfield microscopy. It includes 1,186 training, 1,199 validation, and 201 test images at 540×540 resolution, with up to 105 instances per training image and an average of 96 per test image, making it a challenging real-world benchmark.
QCell outperforms state-of-the-art methods on multiple benchmarks, achieving +2.2 AP and +2.7 AJI on ISBI2014. The paper is accepted at BMVC 2026, and code is available on GitHub.