loihuynh/semisub-cxr-supervised-0.2

Semi-supervised multi-label chest X-ray classifier trained on the CheXpert dataset.

Model Description

• Architecture: ResNet-50 backbone + Dropout + Linear head (6 classes)
• Backbone: resnet50_imagenet (ImageNet pretrained)
• Training setting: supervised
• Labeled data ratio: 0.2 (20% of training set)
• Seed: 42
• Best epoch: 0

Labels

This model predicts the following 6 chest X-ray findings:

Cardiomegaly, Pleural Effusion, Pneumothorax, Consolidation, Atelectasis, Edema

Performance

Macro AUROC: 0.8497

Class	AUROC
Cardiomegaly	0.7625
Pleural Effusion	0.9194
Pneumothorax	0.8520
Consolidation	0.8323
Atelectasis	0.8182
Edema	0.9135

Usage

from src.utils.hub import load_from_hub

model = load_from_hub("loihuynh/semisub-cxr-supervised-0.2")
model.eval()

# Run inference on a preprocessed image tensor (1, 3, 224, 224)
import torch
with torch.no_grad():
    logits = model(image_tensor)
    probabilities = torch.sigmoid(logits)

Training Details

• Dataset: CheXpert
• Method: Supervised baseline
• Loss: Binary Cross-Entropy (with masking for pseudo-labels)
• Image size: 224x224

Citation

If you use this model, please cite the CheXpert dataset:

@inproceedings{irvin2019chexpert,
  title={CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison},
  author={Irvin, Jeremy and others},
  booktitle={AAAI},
  year={2019}
}