XenoScan: time-lapse image dataset of Xenopus laevis embryos under Wnt-pathway compound treatment, with trained model weights

Per-embryo time-lapse images underlying "XenoScan: a deep-learning-based large-scale organismal phenomics analysis platform for aquatic model organism embryos". This dataset contains 186,313 JPEG images of 1,159 individual Xenopus laevis embryos (27.2 GB) with a row-per-image manifest. It is the drug-response classification dataset reported in the paper: 172,090 training images from 1,081 embryos and 14,223 independent test images from 78 embryos. Three treatment groups are included — untreated control, BIO, and Wnt-C59. Contents Thirteen tar archives, one per experiment cycle (C-02…C-14), plus three loose files: README.md, manifest.csv (one row per image) and checksums.md5. Each archive is self-describing, carrying the README together with its own manifest and checksum slices. Cycle C-14 is exactly the independent test set. Image and metadata format RGB JPEG, 512 × 1024 pixels, one embryo per image, cropped from the flatbed scan of its well and rotated to a common orientation. Scale is 1.172 × 10-2 mm/pixel. Metadata is encoded positionally in the path, e.g. images/0036_C-02_P-00_sid-07_01_CONTROL/0036_0000.jpg → embryo id 0036, experiment cycle C-02, plate P-00, well position sid-07, treatment CONTROL, frame 0000. One frame every 25 minutes; developmental time is hpf = frame × 25/60 + 2, spanning 2–84.9 hpf. Composition TreatmentEmbryosImages Control857137,940 Wnt-C5916524,536 BIO13723,837 Total1,159186,313 Developmental windows follow the manuscript's stage definitions: cleavage (frames 0–48), neural (49–67), tailbud (68–128) and free-swimming (129–199). Notes for reuse Join on the label field, never on the two-digit code in the directory name. The code is a legacy field and is not a valid key. Metadata is positional. Renaming or re-nesting an embryo directory destroys it. Experiment cycles are the unit of biological replication and are a batch variable. Treatments are not balanced across cycles: control embryos appear in all 13, BIO in 6, Wnt-C59 in 5. Frame counts vary (median 162). Read frame from the manifest rather than assuming position in a sorted listing. Segmentation masks are not included; they are derived products. Trained model weights XenoScan-models.tar (229 MB) unpacks to a models/ directory holding seven Keras .h5 files — the exact weights behind every result in the paper, trained under TensorFlow/Keras on NVIDIA GeForce RTX 3090 GPUs. A README inside the archive maps each file to its stage. FileModel resnet18_TB-01.h5 … resnet18_TB-04.h5 ResNet-18 drug-response classifiers, one per developmental window (cleavage, neural, tailbud, free-swimming) UNET_XENOPUS_832x512.h5, UNET_XENOPUS_208x128.h5 U-Net body segmentation at two input resolutions; both are applied to every image and their probability maps merged by per-pixel maximum predict_hpf.h5 CNN regressor for developmental time. Its target is frame x 25/60, which is 2 h below the hpf = frame x 25/60 + 2 convention used in manifest.csv and the manuscript. Random seeds were not fixed during training, so retraining will not reproduce these weights bit-for-bit. Analysis code is available at github.com/hmyang-UNIST/XenoScan and is distributed with the paper as Software S1. See README.md for the full description.