---
license: apache-2.0
tags:
- medical
- computer-vision
- image-classification
- multi-label-classification
- radiology
- pytorch
- convnext
- cbam
- timm
datasets:
- stanfordmlgroup/chexpert
metrics:
- auc
pipeline_tag: image-classification
library_name: timm
# THIS LINK ADDS THE DEMO WIDGET TO YOUR MODEL PAGE
demonstrated_in: calender/GRADCAM-Convnext-Chexpert-Attention
---

# ConvNeXt-CheXpert: CBAM-Augmented Thoracic Classifier

[![GitHub](https://img.shields.io/badge/GitHub-Source_Code-black?logo=github)](https://github.com/jikaan/convnext-chexpert-attention)
[![Dataset](https://img.shields.io/badge/Dataset-CheXpert-green)](https://stanfordmlgroup.github.io/competitions/chexpert/)
[![Open In Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Open%20Demo-blue)](https://huggingface.co/spaces/calender/GRADCAM-Convnext-Chexpert-Attention)

## Model Description

This repository contains the weights for a **ConvNeXt-Base** architecture fine-tuned for multi-label classification of chest radiographs. The model is augmented with **Convolutional Block Attention Modules (CBAM)** to enhance feature localization and interpretability.

**Try the live demo here:** [👉 GRADCAM-Convnext-Chexpert-Attention Space](https://huggingface.co/spaces/calender/GRADCAM-Convnext-Chexpert-Attention)

* **Architecture:** ConvNeXt-Base + CBAM
* **Validation AUC:** 0.81 (Iteration 6)
* **Input Resolution:** 384x384

## Detectable Classes

The model outputs probabilities for the following 14 classes:
1. No Finding
2. Enlarged Cardiomediastinum
3. Cardiomegaly
4. Lung Opacity
5. Lung Lesion
6. Edema
7. Consolidation
8. Pneumonia
9. Atelectasis
10. Pneumothorax
11. Pleural Effusion
12. Pleural Other
13. Fracture
14. Support Devices

## Interpretability (Grad-CAM)

The model's attention mechanism allows for precise localization of pathologies.

<div align="center">
  <img src="https://huggingface.co/calender/Convnext-Chexpert-Attention/resolve/main/3.png" width="900" alt="Atelectasis Analysis"/>
  <p><em>Figure 1: Multi-label detection visualized via Grad-CAM.</em></p>
</div>

## Usage

### Prerequisites

```bash
pip install torch torchvision timm pillow
```
Inference Code

This script loads the model and performs inference using the specific normalization statistics from training.
Python
```
import torch
from PIL import Image
from torchvision import transforms
import timm

# 1. Configuration
# Ensure you have downloaded model.pth from this repository
MODEL_PATH = "model.pth" 
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

pathologies = [
    "No Finding", "Enlarged Cardiomediastinum", "Cardiomegaly",
    "Lung Opacity", "Lung Lesion", "Edema", "Consolidation",
    "Pneumonia", "Atelectasis", "Pneumothorax", "Pleural Effusion",
    "Pleural Other", "Fracture", "Support Devices"
]

# 2. Load Model
model = timm.create_model('convnext_base', pretrained=False, num_classes=14)
model.load_state_dict(torch.load(MODEL_PATH, map_location=DEVICE))
model.to(DEVICE)
model.eval()

# 3. Preprocessing
# Note: These specific mean/std values are critical for accuracy
transform = transforms.Compose([
    transforms.Grayscale(num_output_channels=3),
    transforms.Resize((384, 384)),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.5029414296150208] * 3,
        std=[0.2892409563064575] * 3
    )
])

# 4. Run Prediction
image = Image.open('chest_xray.jpg')
input_tensor = transform(image).unsqueeze(0).to(DEVICE)

with torch.no_grad():
    logits = model(input_tensor)
    probs = torch.sigmoid(logits)[0]

print("Predictions:")
for name, score in zip(pathologies, probs):
    print(f"{name}: {score.item():.3f}")
```
Citation

```

@misc{convnext_chexpert_2025,
  author = {Calendar, S.},
  title = {ConvNeXt-CheXpert: CBAM-Augmented Thoracic Classifier},
  year = {2025},
  publisher = {Hugging Face},
  url = {[https://huggingface.co/calender/GRADCAM-Convnext-Chexpert-Attention](https://huggingface.co/calender/GRADCAM-Convnext-Chexpert-Attention)}
}```