How worknestmajor/impression-generation-from-xray-images works
Project type and purpose
• This is a machine-learning application that automates generation of the “impression” section of chest X-ray medical reports.
• It consists of two main phases: a model-training pipeline (create_model.py) and a lightweight inference/serving UI (final.py built on Streamlit).
• It uses chest X-ray images + XML reports (Indiana University dataset) to train a deep learning model and then exposes the model for interactive demo.
File structure
Top-level files and folders:
• README.md ← project description, demo link
• requirements.txt ← Python dependencies
• create_model.py ← data preprocessing & model training script
• final.py ← Streamlit app for inference
• tokenizer.pkl ← saved text tokenizer for report generation
• test_images/1–4 ← sample images to feed into the demo
• __pycache__/ ← Python bytecode cache
Patterns & hints:
• Single-repository, two logical components (train vs. serve)
• No separate “backend” or “frontend” folders, but clear separation by script
• No external databases; data lives in local filesystem
README for additional insights
• Purpose: reduce clinician time by auto-writing report impressions
• Demo hosted via Streamlit badge
• No mention of heavy infra (e.g., Kubernetes), so assume local or simple cloud deployment
Detailed capture of essential elements
• If desired, note external pre-trained backbones (e.g., ResNet) used inside create_model.py.