MicroNucML

Micronuclei are small extranuclear bodies that encapsulate chromosome fragments or whole chromosomes left behind after mitotic errors or unrepaired DNA breaks. They are among the most widely used readouts of chromosomal instability, used to assess cancer risk, screen embryos for aneuploidy, and test compounds for genotoxicity. In practice they are still counted by hand or by intensity thresholding in ImageJ — approaches that either do not scale or require per-experiment parameter tuning, and that readily mistake apoptotic fragmentation for genuine micronuclei.

MicroNucML is a supervised instance-segmentation model that detects and segments micronuclei in immunofluorescence micrographs and then assigns each one to a parent nucleus. It comes from the Harding and Kumar labs at the Princess Margaret Cancer Centre, University Health Network, with collaborators at the University of Toronto and Stanford University, and was posted as a preprint in September 2025. Generalist nuclear segmenters such as Cellpose are trained largely on undamaged nuclei and do not recognize micronuclei; the closest prior tool, micronuclAI, regresses a combined count of micronuclei and nuclear buds rather than producing masks. MicroNucML targets that gap, resolving a structure whose mean segmented area is only 27.2 ± 19.4 pixels at 0.625 µm/pixel.

#Key Features

• Micronucleus-specific instance masks: Returns bounding boxes, centroids, mask areas, RLE masks, and confidence scores per micronucleus rather than a single per-image score, so size and morphology become usable downstream features.
• Parent-nucleus assignment: Nuclei are segmented with SAM2 and each micronucleus is linked to the nearest nucleus by centroid-to-edge distance, enabling metrics such as micronuclei per parent nucleus and the fraction of nuclei that are micronucleus-positive.
• Color and quality invariance: Augmentation over brightness, contrast, saturation, and hue (±30%) plus Gaussian blur makes a model trained only on H2B-GFP images usable on grayscale, red, and blue channels from mCherry constructs or DAPI/Hoechst staining.
• Apoptosis safeguard: A DBSCAN post-processing filter removes dense clusters of five or more detections within 20 pixels, suppressing the blebbed and fragmented nuclei that trip up threshold-based counting.
• Open release: Apache-2.0 code as an installable Python package with a Colab tutorial, checkpoints on HuggingFace, and the 2,137-image annotated corpus under CC-BY on Zenodo.

#Technical Details

MicroNucML is a Mask R-CNN with a ResNet-50 FPN backbone pretrained on ImageNet-1K, fine-tuned with all but the final three layers frozen. The region proposal network was retuned for small objects with anchor sizes of 4, 8, 16, 32, and 64 pixels and a training IoU threshold of 0.5; non-maximum suppression uses an IoU of 0.2, and optimization used AdamW at a learning rate of 0.001. Training labels came from a two-phase pipeline: experts polygon-labeled 148 full-size (1400×1000) images, which were tiled to 224×224 crops to train a ResNet-101 presence classifier (accuracy 0.916, recall 0.991); the classifier then routed likely-positive tiles to brush annotation, whose over-drawn masks were tightened by prompting SAM2. The result is 1,529 training tiles with 2,967 refined micronuclei and 300 held-out tiles with 614, from 687 full-size images of MCF10A and RPE-1 cells under irradiation, replication stress, and cell-cycle stalling.

At a confidence of 0.7 and IoU of 0.5, detection precision, recall, and mAP@50 reach 0.89, 0.91, and 0.96, with inference at 3.3 seconds per image on an NVIDIA T4 and 0.79 seconds on an A100. Mask R-CNN outperformed YOLO, and ResNet-50 beat Swin Transformer and ResNeXt-101 backbones. Because standard IoU penalizes few-pixel errors on very small objects, the authors also report scale-adaptive IoU; under SIOU the F1 score is 0.955 on held-out H2B-GFP images, falling to 0.808 and 0.706 on a lower-quality independent H2B-mCherry set labeled in grayscale and red respectively. Against expert counts on 60 images, MicroNucML tracked manual annotation more closely (β = 1.020, ρ = 0.801) than micronuclAI's CIN score (β = 1.496, ρ = 0.556).

#Applications

The tool suits groups quantifying genome instability at scale: genotoxicity screening, CRISPR and drug screens scored by micronucleus burden, and radiation dose-response studies. In the preprint it was applied to MCF10A cells expressing H2B-GFP and cGAS-mCherry to track cGAS recruitment after 2–20 Gy irradiation, and to ATR inhibition experiments where parent-nucleus assignment revealed a rise in micronuclei per parent nucleus that whole-population normalization obscures. Modeling micronucleus-positive nuclei directly gave a 2.6% increase per gray at 48 hours post-irradiation.

#Impact

MicroNucML supplies both the largest public repository of labeled micronucleus segmentations and a permissively licensed model for a measurement that has been manual for decades, and its SAM2-prompted label refinement is a reusable recipe for building small-object ground truth cheaply. Its limits are acknowledged by the authors: training data come from two human cell lines with H2B-GFP constructs imaged on one Incucyte instrument at 20× and 0.625 µm/pixel, so external datasets need rescaling and segment less cleanly; parent assignment is memoryless and does not track cells through division; and dividing and apoptotic states are excluded rather than classified. The work remains a preprint awaiting peer review.