Deep Learning Instance Segmentation for Wound Healing Assays — Annotated Image Dataset, Trained Models, and Analysis Pipeline

This Zenodo deposit accompanies the manuscript "Deep Learning Instance Segmentation for Quantitative Analysis of Cell Migration in Wound Healing Assays: A Benchmark and Web-Accessible Tool" (submitted to Cytometry Part A, 2026).

CONTENTS:

(1) Image dataset (n = 1,149 brightfield wound-healing assay images):- 815 HUVEC images acquired at 0, 8, 12, and 24 h post-scratch (low-serum conditions, 1% FBS).- 334 SKOV-3 images acquired at 0, 24, 48, and 72 h post-scratch (serum-replete conditions, 10% FBS).- Images provided at 640 x 640 px resolution (resized with black-edge or white-edge padding from 2,452 x 2,056 px native acquisitions) in JPG format. This is the exact representation used by the deployed deep learning models. The higher-resolution native acquisitions are archived separately and remain available from the corresponding author upon request.- COCO Instance Segmentation polygonal annotations (SAM-assisted curation followed by senior expert review) provided per train/valid/test partition (740 / 270 / 139).- Of the 1,149 retained images, 1,020 received a polygonal annotation delineating the wound gap (1,151 polygons total; some images contained multiple non-contiguous gap regions). The remaining 129 images without an assigned polygon correspond to wells in which the wound had reached complete closure, used during training as negative examples.

(2) Trained model weights (4 of the 6 evaluated configurations):- Model 1: YOLOv11 Instance Segmentation, Extra Large, black-edge padding.- Model 2: Roboflow 3.0 Instance Segmentation, Extra Large, black-edge padding (deployed in the web tool).- Model 5: Roboflow 3.0 Instance Segmentation, Extra Large, white-edge padding.- Model 6: YOLOv11 Instance Segmentation, Accurate variant, white-edge padding (deployed in the web tool).

Weights for the two RF-DETR-Seg configurations evaluated in the manuscript (Models 3 and 4) are not included in this deposit and can be regenerated from the deposited dataset following the architecture specifications and training protocol described in the companion paper. Checkpoints are available from the corresponding author upon reasonable request.

All weights provided in PyTorch .pt format. Full training metadata and architectural specifications are provided in models_metadata.json.

(3) Paired ImageJ vs AI measurements (n = 225 observations from 75 unique HUVEC wells x 3 post-scratch time points: 8, 12, and 24 h) used for the method-agreement analysis reported in the companion paper (Pearson r, Lin's Concordance Correlation Coefficient, Bland-Altman with limits of agreement, Two One-Sided Tests for equivalence). Provided as a tidy CSV with annotator, clinical group (EOPE/LOPE), replicate, timepoint, ImageJ closure fraction, and AI closure fraction. Distributed among three specialists (16 wells, 24 wells, and 35 wells respectively), with each well annotated by a single annotator across all three time points.

(4) Analysis pipeline (Python 3.11, SciPy 1.13, statsmodels 0.14, pandas 2.2, matplotlib 3.8) reproducing all statistical results and figures reported in the companion paper.

Live web-deployed inference tool:https://huggingface.co/spaces/nmariotto/Scratch-assay-segmentation

LICENSES:- Image dataset and COCO annotations: Creative Commons Attribution 4.0 International (CC BY 4.0).- Source code (analysis scripts) and trained model weights: MIT License.

See README.md inside the deposit for detailed file structure, citation information, and usage examples.