C2C7 Angle Demo

Demo apps of C2C7 Cobb angle measure by AI presented in our paper (TBA).

Following two versions are available.

• Web app for easy use.
• Executable for more flexible uses.

Web app

Usage

• Open the web app
• Drop a lateral radiograph of cervical spine. Jpeg and PNG input is supported while dicom is not very well tested.
  • Google image search
• Click the measure button.
• Wait for 5 - 30 seconds.

Input	Result

image attribution

Executable

Command line programs.

• c2c7angle: Input one image and output one image.
• c2c7batch: Input a folder/directory, process all images in the input and output the results.

Development

Executable

Executable varsion is coded in rust. Some functions are exported as wasm for the web app version.

Create ONNX model for the web app

The model needs to be compatible with WebGL backend.

1. Use torch.onnx.export to export to ONNX model
2. Apply resize_align_corners.py to change resize mode to align_corners. pytorch_half_pixel is not available in WebGL?
3. Apply onnx-simplifier to optimize away nodes unavailable in WebGL. e.g. python -m onnxsim model.onnx simplefied.onnx --input-shape 2,1,768,768 --dynamic-input-shape

(Note to myself) Do not add the model to the repository. Keep the repo lean.

Weights can be found in google drive.

Tests

Download weights before running tests.

curl -L "https://drive.google.com/uc?export=download&id=1B1EeG38_WTZ42Qg-KlUpnThhEHJQDwkS" -o c2c7.onnx