Blastocyst Grading-Gardner Criteria
It contains the anonymised annotated image corpus, the end-to-end training and evaluation notebook, a demonstration video of the clinical decision-support prototype, and the two reference-standard reliability workbooks that are cited in Section 4.0 (Reference-standard reliability) and Section 4.1 (Cross-model performance) of the manuscript. Clinical context. Day-5 human blastocyst grading on the Gardner (1999) three-parameter system (expansion 1-6, inner-cell-mass A/B/C, trophectoderm A/B/C) is a routine but subjective step in the IVF workflow. Blast-YOLO is a multi-task deep-learning model (ConvNeXt-Tiny backbone with CBAM attention, weighted BiFPN neck, and three parallel task heads for detection, segmentation, and 24-class grading) trained on 19,086 day-5 blastocyst images from a single IVF centre and externally validated on 99 images from an independent centre. The model was released as an embryologist-facing Streamlit prototype that returns the top-three most likely Gardner grades with confidence values. Files in this record. 1. COCO_Segmentation/ — the annotated image corpus in COCO Segmentation format: annotations.json (24-class Gardner labels with per-image polygon segmentations of the embryo) and the accompanying images/ folder. Class definitions follow the 24-class vocabulary defined in Section 3.2 of the manuscript (21 Gardner grades with sufficient training data + three composite clinical-action labels: P1 = High Priority, LP = Low Potential, REV = Review). 2. Blast_Pipeline_v3.ipynb — end-to-end Colab notebook that ingests the COCO folder, converts it to the YOLO layout expected by the Blast-YOLO training loop, trains the model, and evaluates it against the internal test set and external validation set. Reproduces every performance number in Section 4.1 of the manuscript. 3.How_to_use-Blast-YOLO_studio-Streamlit.mp4 — screen-capture demonstration of the seven-mode Streamlit decision-support prototype (Home, Quick Predict, Learn, Proficiency Test, Cohort, Batch, XAI Explorer) that shows the workflow an embryologist follows to grade a single embryo, review the top-three predictions with confidence, generate a per-embryo PDF report, and export batch results. 4. S1_inter-rater_reliability.xlsx — 5 clinical embryologists independently graded the same 100–102 day-5 blastocyst images across two rounds (baseline and after a consensus refresher). Sheets: README, Round_1_baseline, Round_2_post-training, Data_dictionary. 5. S2_intra-rater_reliability.xlsx — the same 5 embryologists independently re-graded 100 day-5 blastocyst images on two separate occasions ≥14 days apart. Sheets: README, Test-retest_all_raters, Data_dictionary.