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" (Cytometry Part A, under revision). This version supersedes v1.0.0 in full. The dataset was rebuilt so that the train/validation/test partition is drawn over acquisition units rather than over individual images, the models were retrained on that partition, and a supervised reference standard was constructed. No file from v1.0.0 is carried over, and no number reported here is comparable to the one it replaces. FILES - dataset.zip (3.7 GB) — brightfield images, the COCO and YOLO annotations, and the partition map that assigns every image both to its split and to the acquisition unit the split was drawn over, so that the absence of leakage can be checked directly from the deposit. Licence: CC BY 4.0.- models_yolo.zip (1.4 GB) — the 25 detection runs, one directory each with best.pt, args.yaml, provenance.json and results.csv. Licence: AGPL-3.0.- models_unet.zip (621 MB) — the 5 runs of the U-Net architectural comparator. Licence: MIT.- reference_standard.zip (67 MB) — supervised reference-standard measurements, ROIs and masks, the agreement tables, the measurement protocol frozen before the campaign (Portuguese record plus a faithful English translation) and a README documenting the layout. Licence: CC BY 4.0.- code.zip (2.8 MB) — the complete pipeline: dataset assembly, training, evaluation, statistical analysis and the deployed application, plus the analysis tables under data/ that the agreement scripts read, provenance CSV/JSON, pinned dependency versions, the pre-registered protocol, the README with installation and reproduction instructions, and the licence texts. Licence: MIT and AGPL-3.0 by file, as set out in NOTICE.- NOTICE — which licence applies to which part, and a correction to the licensing of v1.0.0. DATASET 1,497 brightfield frames were acquired (1,085 HUVEC, 412 SKOV-3). 134 were excluded at a blind quality triage, leaving 1,363. HUVEC were imaged at 0, 8, 12 and 24 h post-scratch under low-serum conditions (1% FBS); SKOV-3 at 0, 24, 48 and 72 h under serum-replete conditions (10% FBS). Images are PNG at the acquired resolution of 2,452 x 2,056 px (1,248 frames); 115 frames are present only as the 640 x 640 export and are labelled as such in the partition map. Annotations are polygonal wound-gap contours, curated with Segment Anything Model assistance followed by expert review: 1,213 images carry 1,350 polygons, some images containing more than one non-contiguous gap. The remaining 150 images have no polygon because the wound had closed, and are retained as negatives rather than discarded. The 1,363 images were grouped into 246 acquisition units — the frames of one well, or one field, followed across time points — and the partition was drawn over those units: train 932, validation 197, held-out test 234. No unit contributes to more than one partition, so no frame of a well used in training appears in the test set. This is the substantive difference from v1.0.0, whose partition was drawn over individual images. MODELS Five configurations were trained at five seeds each, giving 25 runs, under a single-variable design: configuration M (YOLO11m-seg, COCO initialisation, black-edge padding) is the reference, and every other configuration differs from it in exactly one factor — S and X vary the model scale, M-white the padding colour, M-scratch the initialisation. A canonical U-Net was trained under the same partition, schedule and seeds as an architectural comparator. It is not deployed. Weights are PyTorch .pt; each run directory carries its own provenance.json with the package versions, the seed, and evidence of the padding actually present in the training batches. Configurations M and S are the ones served by the web tool, at a confidence threshold of 0.80. REFERENCE STANDARD AND AGREEMENT 223 test frames were measured in ImageJ under parameters frozen in writing before any measurement was taken; that protocol is deposited in reference_standard.zip. A blind visual triage of those frames scored 65 as adequate, 132 as poorly segmented, 14 as invalid images and 12 as ambiguous. After the ambiguous cases were adjudicated and the classification was consolidated with what the annotation and the manual correction subsequently established, the final validity classification is 77 valid, 123 failures of the automatic method and 23 invalid images. The blind triage is kept frozen and reported separately, because it is the only assessment that was not informed by the output it is used to evaluate. 15 frames were re-corrected blind by the same reviewer to quantify intra-observer agreement. The measurements yield 187 closure values, from which the 97 paired observations present in all ten runs (five seeds x two arms), spanning 45 series, enter the agreement analysis against the deployed configuration: Pearson r 0.820, Lin's concordance correlation coefficient 0.803, mean bias +0.053 in the closure fraction. Each statistic is computed within a seed and reported as the mean and standard deviation across the five seeds, so the dispersion quoted is the seed-to-seed variability of the estimator. A frame whose wound had closed completely is recorded as a valid measurement of zero area, not as missing data. The README in reference_standard.zip explains how that is encoded on disk. LICENCES Three apply, by origin, and NOTICE states which covers what. The detection weights and the code that imports ultralytics are AGPL-3.0. Ultralytics distributes YOLO11 under that licence and holds that models trained with it are derivative works, and no commercial licence was obtained. The statistical analysis code, which does not depend on ultralytics, and the U-Net comparator — an independent implementation of the architecture of Ronneberger, Fischer & Brox (2015) — are MIT. The image dataset, its annotations and the reference-standard measurements are CC BY 4.0. v1.0.0 released the detection weights under MIT. That was incorrect for the reason above, and the correction is recorded in NOTICE. Live web tool: https://huggingface.co/spaces/nmariotto/scratch-assay-segmentation