mnDINO

Micronuclei (MN) are small, DNA-containing bodies that form outside the main nucleus when chromosomes or chromosome fragments are mis-segregated during cell division. They are an important biomarker of genome instability, DNA damage, and innate-immune signaling, but they are rare and tiny, which makes them notoriously difficult to detect automatically. General-purpose nucleus segmentation tools routinely miss them, forcing researchers into slow, subjective manual counting.

mnDINO addresses this gap with a vision transformer trained to segment micronuclei in DNA-stained fluorescence images with high accuracy and robustness. Developed in Juan Caicedo's group at the Morgridge Institute for Research (University of Wisconsin–Madison), and described in a March 2026 bioRxiv preprint, the model is built on the DINO self-supervised pretraining strategy. By learning rich image representations without relying solely on scarce hand annotations, mnDINO generalizes across the heterogeneous microscopes, cell lines, and staining conditions that typically break per-dataset segmentation models.

The work comes from a lab with deep roots in bioimage analysis and the CellProfiler ecosystem, and is positioned as a ready-to-use, fixed checkpoint rather than a model that must be retrained for each new microscopy setup.

#Key Features

• Micronuclei-specialized segmentation: Targets the rare, sub-nuclear DNA bodies that general nucleus segmentation tools miss, recovering a biologically important but easily overlooked structure.
• DINO self-supervised backbone: Uses self-distillation pretraining of a vision transformer to learn transferable image features, reducing dependence on large hand-annotated sets.
• Cross-instrument robustness: Trained on a heterogeneous image collection so a single fixed checkpoint generalizes across microscopes and cell lines without per-dataset retraining.
• Fully open release: The dataset, code, and pretrained model are made publicly available to support reproducibility and further micronuclei research.

#Technical Details

mnDINO applies the DINO (self-distillation with no labels) self-supervised framework to train a vision transformer for dense segmentation of micronuclei in DNA-stained fluorescence microscopy. Training uses a heterogeneous, manually curated corpus of more than 5,000 annotated micronuclei spanning multiple cell lines and imaging systems, which is the basis for the model's reported robustness to domain shift. The authors evaluate accuracy and generalization across microscopes and cell lines and show that the single pretrained checkpoint outperforms prior approaches without instrument-specific fine-tuning. The model, annotated dataset, and code are released publicly.

#Applications

mnDINO is aimed at cell biologists, toxicologists, and genome-stability researchers who quantify micronuclei as a readout of DNA damage, chromosomal instability, or cGAS–STING innate-immune activation. Typical workflows include high-content screening of genotoxic compounds, characterizing chromosomal instability in cancer cell models, and studying micronuclei rupture and downstream signaling. Because it ships as a fixed pretrained model, labs can apply it directly to their own DNA-stained images and integrate the segmentation masks into existing image-analysis pipelines.

#Impact

By providing a robust, off-the-shelf segmentation model plus an annotated benchmark dataset, mnDINO standardizes a measurement that has historically relied on manual counting or fragile custom scripts. This lowers the barrier to large-scale, reproducible micronuclei quantification and demonstrates that self-supervised vision transformers can generalize across the messy heterogeneity of real microscopy data. As a recent preprint, its advantages over alternative micronuclei tools await broader independent benchmarking, and—being a fixed checkpoint—performance on imaging modalities far outside the training distribution remains to be established.