XNet38MS is a multi-scale Inception v3 ensemble that takes a chest X-ray image and predicts probabilities for 37 radiological findings. Three models trained at different resolutions (299, 512, 1024 pixels) are averaged at inference time.

Architecture

Backbone: Modified Inception v3 (single-channel input) with 38-class output
Multi-scale: Three models at 299px, 512px, and 1024px, ensembled by averaging sigmoid probabilities
Input: Grayscale chest X-ray image (any resolution, resized and padded internally)

Files

File	Description
`cv/is299/model_best.pth.tar`	Inception v3 trained at 299px (187 MB)
`cv/is299/model_TranslatorCVLogitsToUrgency_fcs.pth.tar`	Urgency head (299px)
`cv/is512/model_best.pth.tar`	Inception v3 trained at 512px (187 MB)
`cv/is512/model_TranslatorCVLogitsToUrgency_fcs.pth.tar`	Urgency head (512px)
`cv/is1024/model_best.pth.tar`	Inception v3 trained at 1024px (187 MB)
`cv/is1024/model_TranslatorCVLogitsToUrgency_fcs.pth.tar`	Urgency head (1024px)

Usage

Download weights

python

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="dnamodel/xraydar-cv",
    local_dir="./xraydar-cv-weights"
)

Place weights for the code repository

python

from huggingface_hub import hf_hub_download
import shutil, os

for size in [299, 512, 1024]:
    dest = f"src/model_20210820_XNet38MS/model_weights/direct_multi93_is{size}_Rv10_pre00_imagenet"
    os.makedirs(dest, exist_ok=True)
    for f in ["model_best.pth.tar", "model_TranslatorCVLogitsToUrgency_fcs.pth.tar"]:
        path = hf_hub_download("dnamodel/xraydar-cv", f"cv/is{size}/{f}")
        shutil.copy(path, os.path.join(dest, f))

See the code repository for full inference instructions.

Radiological Findings (37 classes + 1 meta-class)

#	Finding	#	Finding
0	Abnormal (non-clinically important)	19	Interstitial shadowing
1	Aortic calcification	20	Mediastinum displaced
2	Apical changes	21	Mediastinum widened
3	Atelectasis	22	Object
4	Axillary abnormality	23	Paraspinal mass
5	Bronchial changes	24	Paratracheal/hilar enlargement
6	Bulla	25	Parenchymal lesion
7	Cardiomegaly	26	Pleural abnormality
8	Cavity	27	Pleural effusion
9	Clavicle fracture	28	Pneumomediastinum
10	Consolidation	29	Pneumoperitoneum
11	Cardiac calcification	30	Pneumothorax
12	Dextrocardia	31	Rib fracture
13	Dilated bowel	32	Rib lesion
14	Emphysema	33	Scoliosis
15	Ground-glass opacification	34	Subcutaneous emphysema
16	Hemidiaphragm elevated	35	Tortuosity of aorta
17	Hernia	36	Pulmonary blood flow redistribution
18	Hyperexpanded lungs	37	Volume loss

Citation

bibtex

@article{cid2024development,
  title={Development and validation of open-source deep neural networks for comprehensive chest x-ray reading: a retrospective, multicentre study},
  author={Cid, Yan Digilov and Macpherson, Matt and others},
  journal={The Lancet Digital Health},
  volume={6},
  number={1},
  pages={e44--e57},
  year={2024},
  publisher={Elsevier},
  doi={10.1016/S2589-7500(23)00218-2}
}

License

For academic research and non-commercial evaluation only. See x-raydar.info for terms and conditions.

Contact

Questions or collaborations: Giovanni Montana — g.montana@warwick.ac.uk
Commercial licensing: Warwick Ventures — ventures@warwick.ac.uk