raeeshan2/cxr-report-generator-ckpt3

Chest X-ray draft report generator (research prototype)

Not a medical device and not validated for clinical use. Generated reports can contain invented or missed findings and must be reviewed by a qualified radiologist.

ViT encoder + DistilGPT-2 decoder, with a 12-finding classifier whose output is written as a "findings:" line the decoder continues from. Trained on about 3,000 frontal chest X-rays from the Indiana University (Open-i) collection, using keyword-derived labels.

Files

Besides the model weights, aux_head.pt, meta.json and thresholds_v2.json are needed. Loading requires the code in the project's GitHub repository (cxr_report_generator.py), not plain AutoModel.

Limitations

• Single small dataset; may not generalise to other scanners or populations.
• Frontal chest views only; it does not reject other images.
• Cannot reliably detect consolidation, pneumothorax, pneumonia, nodule/mass, fracture or scoliosis.
• Can describe things that are not in the image.
• Evaluated only with keyword-label metrics, not by radiologists.

Intended use

Research and education only. Not for diagnosis or patient care.