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. Two scales are provided: M.pt is the default and S.pt is a faster alternative that trades about four percentage points of recall for roughly half the CPU latency.

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

⚠️ These are not the weights of the originally submitted version

The first submission evaluated models trained on a dataset partitioned at the image level, so frames of the same acquisition field appeared in both training and test sets. Every performance figure in that version was optimistic. The dataset was rebuilt and partitioned at the level of the physical acquisition field, and all configurations were retrained. The weights here are from the corrected partition. The superseded weights remain available in the earlier deposit, marked as such; they should not be used.

Files

File	Scale	Parameters	Size	CPU (ms/img)	GPU (ms/img)	mAP@50	Recall
M.pt	YOLO11m-seg	22.4 M	45.2 MB	345	92	93.40 ± 1.08	78.34 ± 2.99
S.pt	YOLO11s-seg	10.1 M	20.5 MB	174	85	93.98 ± 0.74	74.32 ± 2.30

Accuracy is mean ± SD over five training seeds on a held-out test set of 234 images from 37 acquisition groups. Latency is the median over 40 images, CPU on 16 cores, GPU on an NVIDIA GeForce RTX 4060 Ti. Both weights are the seed-42 checkpoint.

The two scales are not statistically distinguishable in mAP. Across all five configurations evaluated, mean mAP@50 spans 93.3–94.0% and none of the ten pairwise differences has a cluster-bootstrap confidence interval excluding zero. The reason to prefer M is recall at the deployed operating point, not accuracy in aggregate.

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
Seeds	42–46; the checkpoint published here is seed 42

Training data: 932 images (197 validation, 234 held-out test) of HUVEC and SKOV-3 monolayers, brightfield, 5× objective, single laboratory. One class, wound.

Usage

from ultralytics import YOLO

model = YOLO("M.pt")                  # or S.pt for the fast variant
r = model.predict("scratch.png", conf=0.80, retina_masks=True)
mask = r[0].masks                     # polygon and binary mask of the gap

conf=0.80 is the operating point at which the reported precision and recall were measured and the default of the companion web interface.

What these weights are good for, and where they fail

Agreement with a supervised reference standard, over 97 paired observations from both cell lines: Pearson r 0.820 ± 0.042, Lin's concordance correlation coefficient 0.803 ± 0.049, mean bias +0.053 ± 0.018 in the closure fraction, 95% limits of agreement −0.288 to +0.395.

Read that last number carefully. A single automated measurement can differ from a careful manual one by about thirty percentage points of closure. The workflow is suitable for comparing conditions across many wells; it is not a substitute for manual measurement of an individual well.

Two known limits:

• Small wounds. Performance is governed by wound size rather than elapsed time. Above 10% of the field the model agrees closely with the reference standard; between 2% and 5% mean intersection over union falls to 0.362. Measurements taken while the residual gap is below roughly 5% of the field are the least reliable part of a series.
• 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.

The ceiling is not the architecture. On blinded repeat corrections the human observer reproduced their own delineation at a median intersection over union of 0.861 — about what the models achieve. What limits agreement here is the reproducibility of the wound boundary itself.

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 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