alex17cmbs/chexnet-multilabel-14

🫁 DenseNet-121 (CheXNet) - Multi-Label Chest X-ray Classification (14 Pathologies)

Ce modèle a été entraîné pour la classification multi-label de 14 pathologies thoraciques à partir de radiographies X-ray du dataset ChestX-ray14.

📋 Description

• Architecture: DenseNet-121 (CheXNet)
• Tâche: Classification multi-label (14 pathologies)
• Dataset: NIH Chest X-ray (ChestX-ray14)
• Framework: PyTorch
• Image Size: 224×224

📊 Performance Globale

Métrique	Valeur
AUC-ROC (macro)	0.7486
AUC-ROC (micro)	0.8219
F1 (macro)	0.0472
mAP	0.1908

Comparaison avec l'article CheXNet

Modèle	AUC Macro	Δ
CheXNet (article)	0.8414	-
Notre modèle	0.7486	-0.0928

📈 Performance par Pathologie

Pathologie	AUROC	F1	Support
Atelectasis	0.7216	0.0216	3279
Cardiomegaly	0.8599	0.1680	1069
Effusion	0.7937	0.2974	4658
Infiltration	0.6757	0.0761	6112
Mass	0.7583	0.0747	1748
Nodule	0.6720	0.0061	1623
Pneumonia	0.6577	0.0000	555
Pneumothorax	0.7981	0.0082	2665
Consolidation	0.7023	0.0000	1815
Edema	0.7954	0.0085	925
Emphysema	0.7588	0.0000	1093
Fibrosis	0.7499	0.0000	435
Pleural_Thickening	0.7175	0.0000	1143
Hernia	0.8189	0.0000	86

🏷️ Les 14 Pathologies

ID	Pathologie
0	Atelectasis
1	Cardiomegaly
2	Effusion
3	Infiltration
4	Mass
5	Nodule
6	Pneumonia
7	Pneumothorax
8	Consolidation
9	Edema
10	Emphysema
11	Fibrosis
12	Pleural_Thickening
13	Hernia

⚙️ Configuration d'entraînement

{
  "data_variant": "full",
  "batch_size": 16,
  "image_size": 224,
  "num_classes": 14,
  "learning_rate": 0.001,
  "num_epochs": 50,
  "scheduler": "plateau (factor=0.5, patience=5)",
  "optimizer": "Adam (betas=(0.9, 0.999))",
  "loss": "BCEWithLogitsLoss (non pond\u00e9r\u00e9e)"
}

🚀 Utilisation

import torch
from torchvision import transforms
from PIL import Image

# Charger le modèle
# Pour DenseNet-121
import timm

model = timm.create_model(
    'densenet121',
    pretrained=False,
    num_classes=14
)
model.load_state_dict(torch.load('pytorch_model.bin', map_location='cpu'))
model.eval()

# Préprocessing
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

# Pathologies
PATHOLOGIES = [
    "Atelectasis", "Cardiomegaly", "Effusion", "Infiltration",
    "Mass", "Nodule", "Pneumonia", "Pneumothorax", "Consolidation",
    "Edema", "Emphysema", "Fibrosis", "Pleural_Thickening", "Hernia"
]

# Prédiction
image = Image.open('chest_xray.png').convert('RGB')
input_tensor = transform(image).unsqueeze(0)

with torch.no_grad():
    logits = model(input_tensor)
    probs = torch.sigmoid(logits)

# Afficher les probabilités
for name, prob in zip(PATHOLOGIES, probs[0]):
    print(f"{name}: {prob:.4f}")

📚 Citation

@inproceedings{Wang_2017,
    title = {ChestX-Ray8: Hospital-Scale Chest X-Ray Database and Benchmarks},
    author = {Wang, Xiaosong and Peng, Yifan and Lu, Le and Lu, Zhiyong and Bagheri, Mohammadhadi and Summers, Ronald M},
    booktitle = {IEEE CVPR},
    year = {2017}
}

@article{rajpurkar2017chexnet,
    title={CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning},
    author={Rajpurkar, Pranav and others},
    journal={arXiv preprint arXiv:1711.05225},
    year={2017}
}

📄 License

MIT License