MuscleMap Web App

Browser-based whole-body muscle segmentation using a MONAI 2D UNet model running entirely client-side via ONNX Runtime Web. Segments 99 muscles from MRI — no server required.

Quick Start

# 1. Download ONNX Runtime WASM files
cd web
bash setup.sh

# 2. Start the development server
#    The MuscleMap ONNX models are already committed in web/models/.
bash run.sh

Open http://localhost:8080 in your browser.

Usage

1. Upload a whole-body MRI as DICOM folder, NIfTI image, or NIfTI label map; use Add File(s) to append more NIfTI inputs
2. Assign each loaded NIfTI to its role: T1/T2 SE image, Dixon fat, Dixon water, Dixon opposed-phase, Dixon in-phase, or segmentation label map
3. Choose which image contrasts should run segmentation
4. Optionally adjust sliding window overlap, slice thickness, or enable Low-Res Mode in Inference Settings
5. Click Run Segmentation
6. Inspect each generated segmentation overlay independently in Results
7. In Postprocessing, optionally consolidate multiple segmentations into one label map
8. In Postprocessing, optionally calculate volumetric metrics or IMF metrics. T1/T2 SE metrics use K-means or Gaussian-mixture thresholding from one source image; Dixon metrics require both fat and water images and can be calculated alongside T1/T2 SE IMF.
9. Download segmentation NIfTI label maps or the metrics CSV

All processing happens locally in your browser. Patient data is not uploaded.

Model Conversion

Convert a MuscleMap PyTorch checkpoint to ONNX:

pip install torch monai onnx onnxruntime

# FP32 (full precision, ~54 MB)
python scripts/convert_model.py --checkpoint /path/to/model.pth

# UINT8 quantized (~14 MB, faster download)
python scripts/convert_model.py --checkpoint /path/to/model.pth --quantize

Output is saved to web/models/musclemap-wholebody.onnx.

The ONNX files in web/models/ are committed so the GitHub Pages deployment can serve them directly from the same origin. Temporary PyTorch checkpoints used for conversion are cached in .tmp_weights/, which is ignored by git.

Project Structure

musclemap-webapp/
├── scripts/
│   └── convert_model.py              # PyTorch → ONNX conversion
├── web/
│   ├── index.html                     # Main page
│   ├── setup.sh                       # Downloads ONNX Runtime WASM
│   ├── run.sh                         # Dev server (Python, COOP/COEP headers)
│   ├── css/styles.css
│   ├── js/
│   │   ├── musclemap-app.js           # Main orchestrator
│   │   ├── inference-worker.js        # Web Worker (full inference pipeline)
│   │   ├── app/
│   │   │   ├── config.js              # Model config, inference params
│   │   │   └── labels.js              # 99 muscle labels + colors
│   │   ├── controllers/
│   │   │   ├── FileIOController.js    # NIfTI upload
│   │   │   ├── DicomController.js     # DICOM folder upload + conversion
│   │   │   ├── ViewerController.js    # NiiVue viewer
│   │   │   └── InferenceExecutor.js   # Worker lifecycle
│   │   └── modules/
│   │       ├── inference/
│   │       │   ├── preprocessing.js   # Orient, resample, normalize, crop
│   │       │   ├── sliding-window.js  # 2D sliding window + Gaussian weighting
│   │       │   ├── postprocessing.js  # Label cleanup, inverse transforms
│   │       │   └── connected-components.js
│   │       ├── file-io/
│   │       │   └── NiftiUtils.js
│   │       └── ui/
│   │           ├── ConsoleOutput.js
│   │           ├── ProgressManager.js
│   │           ├── ModalManager.js
│   │           └── MuscleLegend.js    # Detected muscles panel
│   ├── dcm2niix/                      # DICOM→NIfTI WASM module
│   ├── nifti-js/                      # NIfTI parser library
│   ├── wasm/                          # ONNX Runtime WASM (downloaded by setup.sh, ignored by git)
│   └── models/                        # ONNX models committed for GitHub Pages deployment

Inference Pipeline

1. Parse NIfTI (supports gzip, multiple datatypes)
2. Orient to RAS using the affine matrix
3. Resample to 1×1mm in-plane (keep original z spacing)
4. Normalize intensity (z-score over nonzero voxels)
5. Crop foreground bounding box + 20-voxel margin
6. Slice-by-slice 2D inference with sliding window and Gaussian weighting
7. Inverse transform + cleanup — default mode inverse-transforms first and then runs connected components at full output resolution; optional Low-Res Mode cleans in the cropped working volume before inverse transforms
8. Export — write the segmentation back to original space for download and display

Requirements

• A modern browser with WebAssembly support (Chrome, Firefox, Edge, Safari)
• Python 3 (for the development server only)
• For model conversion: torch, monai, onnx, onnxruntime

Acknowledgements

Built on the prostate-fiducial-seg web app template. Uses NiiVue for visualization, ONNX Runtime Web for inference, and MONAI for the model architecture.