CheXFound¶

ViT-Large/16 backbone pretrained with iBOT-style self-supervised learning by RPIDIAL/CheXFound. Implementation lives in the cloned repo; the recipe loads the teacher checkpoint and re-routes the FSDP-shard key layout to flat blocks.IDX.* keys.

Embed dim	Input size	Returns	Extra
1024	512×512	`(transform, encoder)`	`chexfound` + cloned repo + manual checkpoint

Preprocessing: MONAI LoadImage (supports PNG, JPEG, DICOM, NIfTI, …), min-max normalize to uint8, then Resize(512, BICUBIC) → CenterCrop(512) → Normalize(ImageNet).

Install¶

uv pip install -e ".[chexfound]"

Two manual steps:

Clone the upstream repository:

git clone https://github.com/RPIDIAL/CheXFound third_party_models/CheXFound

Download teacher_checkpoint.pth from Google Drive. Place at third_party_models/CheXFound/weights/teacher_checkpoint.pth.

Override default paths¶

Kwarg	Default	Purpose
`chexfound_dir=`	`third_party_models/CheXFound/`	Root of the cloned repo. The recipe adds this to `sys.path` to import the model code.
`checkpoint_path=`	`<chexfound_dir>/weights/teacher_checkpoint.pth`	Explicit path to the teacher `.pth` weight file.

from radharmony.evaluator.backbones import make_chexfound

transform, encoder = make_chexfound(
    chexfound_dir="/shared/repos/CheXFound",
    checkpoint_path="/path/to/models/CheXFound/teacher_checkpoint.pth",
    device="cuda:0",
)

Usage¶

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

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

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

Segmentation mode¶

from radharmony.dataset import SIIMACRPTXTrainDataset

transform, encoder = make_chexfound(device="cuda:0", output_keys={"img", "mask"})
# forward(x) -> Tensor[B, 1024, 32, 32]   (512 / 16 = 32)

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/chexfound_seg/",
)

get_intermediate_layers strips CLS + register tokens; the recipe uses return_class_token=True to retrieve patch tokens for the segmentation path.