ZFScanner: trained models and analysis result tables for flatbed-scanner zebrafish developmental screening

Trained model weights and analysis result tables for the ZFScanner zebrafish developmental-screening pipeline. Together with the well-image dataset (10.5281/zenodo.21868073) and the analysis code (https://github.com/hmyang-UNIST/ZFScanner), these files reproduce every panel reported in the paper. Contents models/ — the five trained networks behind the published results: two U-Nets for body and eye segmentation, the ten-point landmark regressor, and the two ResNet-18 screening models (atad5a and ddb1). csv/ — the result tables, headed by qinfo.csv: twenty phenotypic features per quantified image, with genotype, developmental time and the train/test partition. The per-feature AUCs, LDA and image-based predictions, and the scanner-versus-microscope measurements are here as well. manual-labels/ — the human ground truth: 245 segmentation masks and 409 annotated landmark sets, with the images they were drawn on. blind-predictions/ — occlusion-sensitivity scores with the head, yolk or trunk blacked out. lda-results/ — per-condition feature-based screening tables. Reproducing the figures. Point configs/paths.yaml in the code repository at this archive; every figure except the representative time-lapse is drawn from these tables alone, without the 27 GB image archive and without a GPU. Model weights are included, so the training scripts are optional. See docs/reproduction.md in the repository for the panel-by-panel map.