VP Shunt Valve Detector – YOLOv8‑Large

A fine‑tuned YOLOv8‑large model for automatic detection and classification of 11 ventricular‑peritoneal (VP) shunt valves on skull radiographs.

---

🌟 Key Features

• Ready‑to‑use weights (weights/vpshunt_yolov8l.pt) – just load and run.
• Gradio demo – launch an interactive web UI or Jupyter notebook in seconds.
• YOLO‑style dataset & config – standard Ultralytics layout with data/vpshunt.yaml bundled.
• Extendable – retrain or add classes with only a few lines of code.

---

📂 Repository Layout

VP‑Shunt‑Detector/
├── data/
│   └── vpshunt.yaml              # dataset config (shown below)
├── weights/
│   └── vpshunt_yolov8l.pt        # fine‑tuned model (≈ 90 MB)
├── notebooks/
│   └── inference_gradio.ipynb    # launchable demo
└── README.md                     # you are here

Note: vpshunt.yaml is packaged so that Ultralytics can locate it automatically for both training and inference:

from ultralytics import YOLO
model = YOLO('weights/vpshunt_yolov8l.pt')     # names baked into weights
# ...but keep yaml alongside weights for reproducibility & re‑training

---

🚀 Quick‑start

1. Install dependencies

Follow the official Ultralytics Quickstart instructions: https://docs.ultralytics.com/quickstart/

For a minimal environment (Linux/macOS):

python -m venv .venv && source .venv/bin/activate
pip install --upgrade pip
pip install ultralytics gradio opencv-python

2. Launch the Gradio demo

python - <<'PY'
from ultralytics import YOLO
import gradio as gr
model = YOLO('weights/vpshunt_yolov8l.pt')  # weights autoload class names

def predict(img):
    results = model(img, conf=0.25)
    return results[0].plot()

gr.Interface(fn=predict,
             inputs=gr.Image(type='pil', label='Skull Radiograph'),
             outputs=gr.Image(label='Detections')).launch()
PY

Or open notebook/inference_gradio.ipynb and run the cells.

---

🔧 Fine‑tuning / Adding Classes

1. Prepare your dataset in YOLO format (images/ + .txt labels).
2. **Update ** data/vpshunt.yaml or create a new YAML including all classes (existing + new).
3. Train starting from these weights:

# example: continue training 100 epochs at 640 px
yolo train model=weights/vpshunt_yolov8l.pt \
           data=data/vpshunt.yaml \
           epochs=100 imgsz=640 \
           project=runs_vpshunt name=exp_finetune

YOLO automatically expands the detection head for new classes when the YAML specifies a larger names list.

---

📑 Dataset Config

path: yolo # adapt to your directory
train: images/train
val: images/val
test: images/test

names:
  0: Codman Certas
  1: Medtronic Strata
  2: Codman Hakim
  3: Medtronic Delta
  4: Aesculap proGAV
  5: Sophysa Polaris
  6: Sophysa Sophy
  7: Aesculap proGAV 2
  8: Integra DP
  9: Spitz Holter
  10: Codman Fixed

Adapt path: if your dataset lives elsewhere.

---

⚠️ Disclaimer

This repository is provided as‑is for research and educational purposes. The model has not been evaluated or cleared by the U.S. Food & Drug Administration (FDA) and is NOT intended for clinical decision‑making or patient care. Use at your own risk.

---

📝 License

• Ultralytics YOLOv8 framework: AGPL‑3.0 – see https://github.com/ultralytics/ultralytics/blob/main/LICENSE