RAD-DINO¶

Microsoft's chest-X-ray DINOv2 ViT — microsoft/rad-dino on HuggingFace.

Embed dim	Input size	Returns	Extra
768	518×518	`(transform, encoder)`	`raddino`

ViT-B/14 (14-px patches). Preprocessing follows the RAD-DINO processor: 8-bit grayscale → 3-channel uint8 → resize 518 → ImageNet normalize.

Install¶

uv pip install -e ".[raddino]"

Weights are auto-downloaded from HuggingFace on first call.

Usage¶

from radharmony.evaluator.backbones import make_raddino
from radharmony.dataset import VinDrCXRTrainDataset

transform, encoder = make_raddino(device="cuda:0", output_keys={"img", "cls"})

ds = VinDrCXRTrainDataset(
    base_image_dir="/data/vindr/train/",
    transform=transform,
    cache_dir="/tmp/cache/raddino/",
    output_cls=True,
)

Segmentation mode¶

from radharmony.dataset import SIIMACRPTXTrainDataset

transform, encoder = make_raddino(device="cuda:0", output_keys={"img", "mask"})
# forward(x) -> Tensor[B, 768, 37, 37]   (518 / 14 = 37)

ds = SIIMACRPTXTrainDataset(
    base_image_dir="/data/siim-acr-ptx/dicom-images-train/",
    csv_path="/data/siim-acr-ptx/train-rle.csv",
    transform=transform,
    mask_output_dir="/tmp/cache/siim_ptx_masks/",
    output_mask=True,
    cache_dir="/tmp/cache/raddino_seg/",
)