You can deploy this model and start earning money today!

Beiry/
medscan-chest-xray14

Name: medscan-chest-xray14
Availability: InStock

Views

No views yet

MedScan ChestX-ray14

Multi-label thoracic finding prediction for NIH ChestX-ray14.

This project trains a 15-output chest X-ray classifier for the MedScan Streamlit app. The app expects checkpoints/chest_best.pt; the checkpoint saved by train_chest_multilabel.py includes the model weights, class list, model architecture, input size, thresholds, validation metrics, and training arguments.

Dataset

NIH ChestX-ray14 contains frontal chest X-ray images with labels parsed from radiology reports. The raw images are not included in this repository and must not be uploaded to Hugging Face with the model.

Expected local layout:

text

chest-xray14-medscan/
  Data_Entry_2017_v2020.csv
  train/
    images/
      00000001_000.png
      ...

Task

15-class multi-label chest X-ray classification.

Classes:

text

Atelectasis, Cardiomegaly, Consolidation, Edema, Effusion,
Emphysema, Fibrosis, Hernia, Infiltration, Mass,
No Finding, Nodule, Pleural_Thickening, Pneumonia, Pneumothorax

Model

Default architecture: convnext_base.fb_in22k_ft_in1k from timm.

Default input size: 224 x 224.

Loss: BCEWithLogitsLoss with training-set pos_weight for class imbalance.

Validation metrics saved in runs/chest_multilabel/history.json:

train loss
val loss
micro-F1
macro AUC
per-class AUC

Usage

Smoke test:

.\.conda\python.exe train_chest_multilabel.py --device cuda --epochs 1 --batch-size 4 --num-workers 0 --max-samples 128 --smoke-test

Full training:

.\.conda\python.exe train_chest_multilabel.py --device cuda --epochs 20 --batch-size 32 --num-workers 8

The best app-compatible checkpoint is saved to:

checkpoints/chest_best.pt

Hugging Face Upload

Set a Hugging Face token, then upload the app-compatible model artifacts:

powershell

$env:HF_TOKEN="hf_your_token_here"
.\.conda\python.exe upload_model_to_hf.py --repo-id yourname/medscan-chest-xray14

After upload, the Streamlit app can download the model from the Hub by setting:

$env:MEDSCAN_CHEST_HF_REPO="yourname/medscan-chest-xray14"

The app still prefers the local checkpoints/chest_best.pt when present. If the local file is missing, it downloads checkpoints/chest_best.pt from the configured Hugging Face repo.

Intended Use

For research and educational purposes only. This model is not a medical device and must not be used for clinical diagnosis or treatment decisions.

Beiry/medscan-chest-xray14