Embryo Scan blastocyst grader
Five ConvNeXt-Tiny models (5-fold cross-fitting) that grade a day-5 human blastocyst photo on the Gardner scale: expansion (1-5), inner cell mass (A-C) and trophectoderm (A-C). Code, training and evaluation: github.com/cha205/embryo-scan.
Results
On the 300-photo test set graded by a panel of embryologists (Kromp et al. 2023), the 5-model ensemble with 8-view test-time augmentation scores accuracy 0.853 (expansion), 0.714 (inner cell mass) and 0.660 (trophectoderm), against 0.78 / 0.74 / 0.70 for individual embryologists and 0.85 / 0.69 / 0.62 for the best model published with the dataset. A check that retrains without the test set's patients finds no sign that patient overlap inflates these numbers.
Files
fold-0.pt ... fold-4.pt: PyTorch state dicts for embryoscan.blastocyst.grading.Grader (timm convnext_tiny.fb_in22k_ft_in1k backbone, three heads). Input: grayscale 384 x 288. Load them with embryoscan.blastocyst.inference.load_ensemble.
Training data and license
Trained on the blastocyst dataset of Kromp et al. 2023, An annotated human blastocyst dataset to benchmark deep learning architectures for in vitro fertilization, Scientific Data (figshare, CC BY 4.0). The weights are released under CC BY 4.0.
Limitations
- Photos come from one clinic and one microscope; other setups may grade poorly.
- Grades follow a single embryologist's training labels, which tend to call grade-B inner cell masses "A".
- Research project, not a medical device.