ReconstructUS

ReconstructUS is an end-to-end pipeline for ultrasound image reconstruction from raw RF data.
It combines classical beamforming techniques, deep learning models, and FPGA-based deployment.
This project serves as a proof-of-concept showing that high-quality ultrasound image reconstruction from a single plane wave is possible.

---

Overview

The pipeline integrates:

• MATLAB preprocessing: Convert .uff files from the PICMUS/USTB datasets into .mat format using the Ultrasound Toolbox and perform Delay-And-Sum (DAS) reconstruction.
• Python deep learning training: Train U-Net and EfficientUNet models for ultrasound image reconstruction.
• FPGA deployment: Quantize, compile, and deploy trained models onto the Xilinx Kria KV260 Vision AI Starter Kit using the Vitis AI toolchain.

---

Project Workflow

1. Dataset acquisition
   Download in vivo, experimental, and simulation datasets from the USTB dataset collection.

2. Preprocessing
   Use MATLAB scripts in matlab/ to process .uff files into .mat files.

   • The .mat files generated during preprocessing are not included in this repository.
   • You must generate them locally using the provided scripts and the original datasets.

3. Model training
   Use the provided Jupyter notebooks:

   • train_unet.ipynb
   • train_effunet.ipynb
     These can be run in Google Colab. They use helper functions from:
   • dataloader/ (data loading utilities)
   • models/ (model definitions)
   • utils/ (training and evaluation utilities)

4. FPGA deployment
   Follow the vitis-ai/ directory for the quantization, compilation, and deployment pipeline.
   Detailed instructions are in vitis-ai/HOWTO.md.

---

Repository Structure

dataloader/

Data loading utilities for training and evaluation.

matlab/

MATLAB scripts for preprocessing .uff files from the USTB datasets into .mat files and performing Delay-And-Sum (DAS) reconstruction.

models/

Model definitions, including U-Net and EfficientNet-based U-Net.

results/

Example output images from final reconstructions.

utils/

Helper functions for training, evaluation, and metric computation.

vitis-ai/

Full Vitis AI deployment pipeline for the Kria KV260:

• Quantization
• Compilation
• Deployment Includes HOWTO.md with step-by-step deployment instructions.

train_effunet.ipynb

Google Colab-ready notebook to train the EfficientNet-based U-Net.

train_unet.ipynb

Google Colab-ready notebook to train the U-Net.

requirements.txt

Python dependencies for training and evaluation.

---

Requirements

See requirements.txt for Python dependencies.
For FPGA deployment, additional dependencies (such as xir) must be installed inside the Vitis AI Docker container — do not install them locally.

---

License

This project is licensed under the MIT License – see the LICENSE file for details.

---

Final Notes

If you find this repository useful, please consider starring it on GitHub or mentioning it in your work.
At some point in the future, I may add a Linux-based workflow for Vitis AI, but for now this repository is complete.

I hope it helps you either reproduce my results or inspires you to create something even better.

Happy computing!

---