cura-chest-xray

Cura: Chest X-Ray Pathology Detection Ensemble

An ensemble of three fine-tuned CNNs for multi-label classification of 14 chest pathologies from frontal chest X-rays.

Model Description

Cura is an end-to-end medical imaging pipeline that performs automated analysis of chest X-rays. This repository contains the three-model ensemble (DenseNet121 + DenseNet169 + ConvNeXt-Tiny) that achieved mean AUC of 0.8539 on the resized NIH ChestX-ray14 test set images (224x224), surpassing the original CheXNet paper (0.841).

Performance

Ensemble Results (NIH ChestX-ray14 Test Set)

Architecture Comparison

Training Details

• Dataset: NIH ChestX-ray14 — 83,703 images (224×224)
• Split: Patient-aware (24,644 train / 3,080 val / 3,081 test patients)
• Loss: AsymmetricLoss (gamma_neg=4, gamma_pos=1, clip=0.05)
• Optimizer: AdamW (lr=1e-4, weight_decay=1e-4)
• Scheduler: CosineAnnealingLR (T_max=20, eta_min=1e-6)
• Batch size: 32 (DenseNet121, ResNet50, EfficientNet-B0)
• Batch size: 64 (DenseNet169, ConvNeXt-Tiny)
• Normalization: Dataset-specific (mean=0.4967, std=0.2478)
• Augmentation: RandomRotation(7°), RandomAffine(translate=0.08), ColorJitter(brightness=0.1, contrast=0.1)
• Training platform: Kaggle (single NVIDIA T4 GPU, AMP enabled)

Disease Labels

DISEASES = [
    'Atelectasis', 'Cardiomegaly', 'Effusion', 'Infiltration',
    'Mass', 'Nodule', 'Pneumonia', 'Pleural_Thickening',
    'Pneumothorax', 'Consolidation', 'Edema', 'Emphysema',
    'Fibrosis', 'Hernia'
]

Usage

from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
import torch
import torchvision.models as models

REPO_ID = "mjrq/cura-chest-xray"

def load_model(architecture, device):
    path = hf_hub_download(repo_id=REPO_ID, 
                           filename=f"{architecture}_best.safetensors")
    
    if architecture == 'densenet121':
        model = models.densenet121(weights=None)
        model.classifier = torch.nn.Sequential(
            torch.nn.Dropout(0.2),
            torch.nn.Linear(model.classifier.in_features, 14)
        )
    elif architecture == 'densenet169':
        model = models.densenet169(weights=None)
        model.classifier = torch.nn.Sequential(
            torch.nn.Dropout(0.2),
            torch.nn.Linear(model.classifier.in_features, 14)
        )
    elif architecture == 'convnext_tiny':
        model = models.convnext_tiny(weights=None)
        model.classifier[2] = torch.nn.Sequential(
            torch.nn.Dropout(0.2),
            torch.nn.Linear(model.classifier[2].in_features, 14)
        )
    
    state_dict = load_file(path, device=str(device))
    model.load_state_dict(state_dict)
    model.eval()
    return model.to(device)

Intended Use

This model is intended as a clinical decision support tool. It is meant to be an aid to radiologists, not a replacement. It should not be used as the sole basis for clinical decisions.

Limitations

• Trained on resized NIH ChestX-ray14. Performance may vary on images from different equipment, populations, or imaging protocols.
• Label noise in NIH dataset (~10-20% estimated error rate) limits achievable F1 scores.
• Not validated on non-US patient populations.

Known Limitations by Disease

• Infiltration (AUC 0.72): Lowest performing disease due to non-specific, diffuse appearance overlapping with many conditions.
• Pneumonia (F1 0.11): Poor F1 despite reasonable AUC which reflects severe class imbalance and label noise in NIH dataset.

Citation

If you use this model, please cite:

@software{cura2026,
  author = {mjrq},
  title = {Cura: An AI-Powered Application for Chest X-Ray Analysis},
  year = {2026},
  url = {https://github.com/mjrq/Cura}
}

References

• Rajpurkar et al. (2017). CheXNet. arXiv:1711.05225
• Wang et al. (2017). ChestX-ray8. CVPR 2017
• Ridnik et al. (2021). Asymmetric Loss. ICCV 2021
• Cohen et al. (2022). TorchXRayVision. MIDL 2022
• Resized Dataset: Kaggle — NIH Chest X-ray 14 (224x224 resized)