ZFScanner: a flatbed-scanner well-ROI image dataset of zebrafish embryonic development (72-141 hpf) for ddb1-/-, atad5a-/- and wild-type embryos

Per-well regions of interest cut from office flatbed-scanner images of 96-well plates, covering zebrafish (Danio rerio) development from 72 to 141 hours post-fertilization at 3-hour intervals with six replicate scans per time point. 220,416 images, 1,536 wells, 16 plates, 2,296 plate scans. Each image is 704 x 704 px RGB at 2400 DPI (10.58 um/pixel). The archive is complete at the plate level: every well of every scan is included, and nothing is filtered out - empty wells, motion-blurred acquisitions and non-ventral views are all present, so the quality-control stage of the analysis can be reproduced rather than inherited. The accompanying manifest.csv gives genotype, developmental time and embryo identity for every image, and flags the 95,221 images that passed quality control — a yield of 43.2 % — and, separately, the 81,399 of those that entered the screening analysis. Fifteen of the sixteen plates contributed to the published analysis. EXP-W-01 is a wild-type-only plate whose images pass quality control but were not used to train or evaluate the final classifier — the wild-type embryos in the screening are the siblings sharing each mutant plate — so it carries passed_qc = 1 and used_in_screening = 0. Genotypes: wild type (137,064 images / 956 wells), ddb1-/- (19,440 / 135), atad5a-/- (16,824 / 118), inconclusive genotyping (2,160 / 15), empty wells (44,928 / 312). Files are one ZIP per plate; unpacking them side by side rebuilds the tree that manifest.csv indexes. See README.md for the file-naming scheme, manifest columns, per-plate inventory and image-format notes. Code and models. The analysis code that produced every published result is at https://github.com/hmyang-UNIST/ZFScanner. The trained model weights and the analysis result tables are deposited separately at 10.5281/zenodo.21895906, so the figures can be reproduced from the tables alone without downloading this 27 GB archive.