---
license: mit
library_name: transformers

pipeline_tag: image-classification

base_model: facebook/convnextv2-tiny-22k-224
tags:
  - chest-xray
  - convnext
  - convnextv2
  - image-classification
  - medical
  - medical-imaging
  - pneumonia
  - pytorch
  - transformers
datasets:
  - keremberke/chest-xray-classification

metrics:
  - accuracy
  - f1

widget:
  - src: https://huggingface.co/kiselyovd/chest-xray-classifier/resolve/main/samples/bacterial_pneumonia.png
    example_title: bacterial_pneumonia
  - src: https://huggingface.co/kiselyovd/chest-xray-classifier/resolve/main/samples/normal.png
    example_title: normal
  - src: https://huggingface.co/kiselyovd/chest-xray-classifier/resolve/main/samples/viral_pneumonia.png
    example_title: viral_pneumonia


model-index:
  - name: kiselyovd/chest-xray-classifier
    results:
      - task:
          type: image-classification
        dataset:
          type: keremberke/chest-xray-classification
          name: Chest X-Ray Images (Pneumonia)
        metrics:
          - type: auroc_macro_ovr
            value: 0.9752638346619307
          - type: accuracy
            value: 0.9134615384615384
          - type: macro_f1
            value: 0.9029730638714358

---

# chest-xray-classifier

<p align="center">
  <img src="hero.jpg" width="100%" alt="chest-xray-classifier - 3-class pneumonia classification" />
</p>

A 3-class chest X-ray classifier that distinguishes **bacterial pneumonia**, **normal**, and **viral pneumonia** on pediatric frontal chest radiographs. The main model is a [ConvNeXt-V2-Tiny](https://huggingface.co/facebook/convnextv2-tiny-22k-224) fine-tuned with the `transformers` library; a DINOv2 ViT-S linear probe serves as the baseline.

> Medical disclaimer: this model is provided for research and educational purposes only. It is not a medical device and must not be used for clinical diagnosis or treatment decisions.

## Metrics

Evaluated on the held-out test split (624 radiographs: 242 bacterial / 234 normal / 148 viral). All values are reported as percentages.

| Model | Accuracy | Macro F1 | Macro AUROC (OvR) |
|---|---|---|---|
| **ConvNeXt-V2-Tiny** (main) | **91.3%** | **90.3%** | **97.5%** |
| DINOv2 ViT-S linear probe (baseline) | 85.6% | 84.2% | 94.2% |

### Per-class results (main model)

| Class | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| bacterial_pneumonia | 89.1% | 98.3% | 93.5% | 242 |
| normal | 99.5% | 89.7% | 94.4% | 234 |
| viral_pneumonia | 83.6% | 82.4% | 83.0% | 148 |
| **Macro avg** | **90.7%** | **90.2%** | **90.3%** | 624 |
| **Weighted avg** | **91.7%** | **91.3%** | **91.3%** | 624 |

## Visualizations

All charts below are generated by [`scripts/make_plots.py`](https://github.com/kiselyovd/chest-xray-classifier/blob/main/scripts/make_plots.py), which runs this exact model over the full test split. The computed accuracy / macro-F1 / macro-AUROC are cross-checked against the committed evaluation report before any chart is rendered.

### Confusion matrix

![Confusion matrix](https://huggingface.co/kiselyovd/chest-xray-classifier/resolve/main/plots/confusion_matrix.png)

Counts and row-normalized rates on the 624-image test split. Bacterial pneumonia and normal are recognized strongly; the main source of error is viral pneumonia being predicted as bacterial pneumonia, which is the clinically hardest distinction in this dataset.

### ROC curves (one-vs-rest)

![ROC curves](https://huggingface.co/kiselyovd/chest-xray-classifier/resolve/main/plots/roc_curves.png)

One-vs-rest ROC curves with per-class AUROC. The normal class separates near-perfectly, while viral pneumonia is the most challenging, consistent with the confusion matrix.

### Sample predictions

![Sample predictions](https://huggingface.co/kiselyovd/chest-xray-classifier/resolve/main/plots/sample_predictions.png)

Representative correctly-classified test radiographs (two per class) with predicted class and softmax confidence.

## Usage

```python
import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModelForImageClassification

processor = AutoImageProcessor.from_pretrained("kiselyovd/chest-xray-classifier")
model = AutoModelForImageClassification.from_pretrained("kiselyovd/chest-xray-classifier")
model.eval()

image = Image.open("your_radiograph.png").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits

probs = torch.softmax(logits, dim=-1)[0]
pred_id = int(probs.argmax())
print(model.config.id2label[pred_id], float(probs[pred_id]))
```

Class index order (model `id2label`): `0 -> bacterial_pneumonia`, `1 -> normal`, `2 -> viral_pneumonia`.

## Training data

Trained on the [Chest X-Ray Images (Pneumonia)](https://huggingface.co/datasets/keremberke/chest-xray-classification) dataset of pediatric frontal chest radiographs. The original two-class layout (`NORMAL` / `PNEUMONIA`) is split into three classes by separating the `PNEUMONIA` images into bacterial and viral cases, producing the `bacterial_pneumonia` / `normal` / `viral_pneumonia` target space.

## Intended use and limitations

Research and educational use only. This model is not a medical device, has not been clinically validated, and must not be used for diagnosis or any clinical decision-making. It was trained on a single pediatric dataset and may not generalize to adult radiographs, other imaging equipment, or different acquisition protocols. Users are responsible for evaluating fitness for their use case, including fairness, safety, and compliance with applicable regulations.

## Source code

[GitHub Repository](https://github.com/kiselyovd/chest-xray-classifier)

## License

MIT.