nmariotto/scratch-assay-segmentation

Scratch-assay wound segmentation — YOLO11-seg

Instance-segmentation weights for the cell-free gap in brightfield scratch (wound-healing) assay images. M_v3.pt is the current model and the one the companion web interface serves.

These weights accompany the manuscript "Deep Learning Instance Segmentation for Quantitative Analysis of Cell Migration in Wound Healing Assays" (Cytometry Part A, under revision).

Files

File	Version	Status
M_v3.pt	v3.0.0	current: trained on the well-level partition; evaluated below
M.pt, S.pt	v2.0.0	superseded, kept for the record

Why v2.0.0 was superseded. In v2.0.0 the SKOV-3 images were partitioned by acquisition field, and fields of the same well could fall in different partitions. The field labels were also wrong for most SKOV-3 images: they had been numbered from the first image of each block present in the dataset, so a missing image shifted every label after it. The labels were rederived from the acquisition order of the laboratory's original files, the dataset was partitioned at the level of the physical well, and every configuration was retrained. The weights trained on the original image-level partition stay in the earlier deposit (10.5281/zenodo.20298130), which is marked as superseded and should not be cited.

M_v3.pt

Architecture	YOLO11m-seg, 22.4 M parameters, 45.2 MB
Training seed	42 (one of five; the others are in the data archive)
Default confidence	0.35, chosen on the validation images as the threshold with the smallest mean absolute error in wound area against the annotations
Latency	median over 40 test images at confidence 0.35, taken as the median of three runs: 270 ms on a 16-core CPU, 67 ms on an NVIDIA GeForce RTX 4060 Ti

On the 247 held-out test images (from 25 physical units, none of them present in training):

	at conf 0.35	at conf 0.80
mAP@50 / mAP@50–95	95.9% / 81.7%	(threshold-independent)
Instance precision / recall	86.4% / 94.5%	98.0% / 81.8%
Images with a wound in which one is detected	99.5%
Images without a wound in which one is detected	2 of 27

Closure-fraction agreement with a supervised manual reference standard, at conf 0.35, over 116 paired observations from all 50 eligible acquisition series (a usable baseline in every series): concordance correlation coefficient 0.869, mean bias −0.009, 95% limits of agreement −0.272 to +0.255. In 2 of the 116 observations the model returned no wound where the reference still measured one (a false complete closure).

Training

Architecture	YOLO11-seg (Ultralytics 8.4.102)
Initialisation	COCO checkpoint
Input	640 × 640, letterbox with black padding
Schedule	100 epochs, batch 4, early stopping disabled
Augmentation	HSV 0.015/0.7/0.4, translate 0.1, scale 0.5, fliplr 0.5, mosaic (off for the last 10 epochs)
Framework	PyTorch 2.6.0, CUDA 12.4

Training data: 903 images (213 validation, 247 held-out test) of endothelial cells (EA.hy926) and SKOV-3 monolayers, brightfield, 5× objective, single laboratory, partitioned at the level of the physical well. One class, wound.

Usage

from ultralytics import YOLO

model = YOLO("M_v3.pt")
r = model.predict("scratch.png", conf=0.35, retina_masks=True)
mask = r[0].masks                     # polygon and binary mask of the gap

An empty result (r[0].masks is None) means that no wound was detected. It is not evidence of complete closure, and should be inspected rather than recorded as an area of zero.

What these weights are good for, and where they fail

A single automated measurement can differ from a careful manual one by more than twenty percentage points of closure. These weights measure wound area under conditions comparable to those of the study. They were not validated for comparing treatments: on the held-out data the slope of this model's closure on the reference's, within time point, was 0.76, so a difference between conditions measured with them would be, on average, about a quarter smaller than the manual one.

Known limits:

• Small wounds. Between 2.5% and 5% of the field the mean intersection over union with the annotation is 0.81, against 0.92 above 10%. Measurements taken while the residual gap is below roughly 5% of the field are the least reliable part of a series.
• A stricter threshold misses wounds. At conf 0.80 the model detected a wound in about half of the test images whose wound occupies 2.5–5% of the field, and missed some large baseline wounds as well; an undetected wound reads as a closed one.
• Isolated cells. The gap is quantified as the region enclosed by the segmented contour, so cells that detach and migrate individually into an otherwise continuous gap are not subtracted from it.

Acquisition envelope: one inverted microscope, brightfield, 5× objective, one institution. No external testing was performed. Robustness to other microscopes, magnifications or contrast modalities is unknown.

Licence

AGPL-3.0. These weights are trained with Ultralytics YOLO11, which is AGPL-3.0, and no commercial licence was obtained; the trained weights are a derivative work and carry the same licence. The image dataset is released separately under CC BY 4.0, and the statistical analysis code under MIT.

Citation

Data and code archive: https://doi.org/10.5281/zenodo.20298129 (concept DOI, always the latest). M_v3.pt belongs to v3.0.0 of the archive, https://doi.org/10.5281/zenodo.23453791. Code repository: https://github.com/nykemariotto/scratch-assay-segmentation Web interface: https://huggingface.co/spaces/nmariotto/Scratch-assay-segmentation

Contact: Allan F. F. Alves — allan.alves@unesp.br — ORCID 0000-0002-0954-9919