koscielnamarta/synthetic-usg-afast-vet

Synthetic Veterinary Ultrasound Dataset -- AFAST/POCUS (Dogs & Cats)

Procedurally generated synthetic ultrasound images for 4-class AFAST window classification (CC, DH, HR, SR) in small-animal veterinary medicine.

NOT real ultrasound. NOT clinically usable. Generator: numpy/PIL with Rayleigh speckle noise + geometric anatomical primitives. Intended for pipeline sanity checks and ML architecture testing.

Splits

Split	Images	Per class	Seed	Purpose
train	3200	800	42	Training
test	800	200	42	Validation / early stopping / ablation
holdout	800	200	123	Final hold-out evaluation only (model never seen during training)

Methodological note: holdout uses a different generator seed (123 vs 42 for train/test) to provide an honest final evaluation set. Model selection and hyperparameter tuning use test split; holdout is held back until the very end of the experiment.

Classes (AFAST windows)

Code	Window
CC	Cysto-Colic
DH	Diaphragmatico-Hepatic
HR	Hepato-Renal
SR	Spleno-Renal

How to Load

from huggingface_hub import snapshot_download
local_dir = snapshot_download(repo_id="koscielnamarta/synthetic-usg-afast-vet", repo_type="dataset")

# Train / val / final test
from torchvision.datasets import ImageFolder
train_ds   = ImageFolder(f"{local_dir}/dataset/train")
val_ds     = ImageFolder(f"{local_dir}/dataset/test")
holdout_ds = ImageFolder(f"{local_dir}/dataset/holdout")

Regulatory Note (EU AI Act)

Model fine-tuned exclusively on this synthetic veterinary dataset is not classified as high-risk AI under EU AI Act Art. 6 + Annex III. Production deployment would require fine-tuning on real veterinary ultrasound data and a new regulatory assessment.

Source Code

Generator: https://github.com/koscielnamarta/vet-eye-ai-usg-demo