Published December 30, 2025 | Version zenodo3
Software Open

UNET Acceleration on FPGA

Authors/Creators

  • 1. ROR icon University of Ostrava

Contributors

Hosting institution:

  • 1. ROR icon University of Ostrava

Description

UNET FPGA Acceleration

Affiliation: University of Ostrava, Faculty of Medicine

This solution is designed for the diagnostic analysis of Parkinson’s Disease using transcranial ultrasound imaging. The pipeline operates in two primary stages:

  1. Brain Stem Localization: The first deep neural network (U-Net) processes the raw ultrasound data to identify and segment the brain stem mask.

  2. Substantia Nigra Identification: This segmented region is then used as the Region of Interest (ROI) for a second specialized network. Its goal is to identify the Substantia Nigra (SN)—the primary focus of this research, as its echogenicity is a key biomarker for Parkinson’s.

Final Output: Once the objective is identified, the system automatically fits an elliptical regressor around the Substantia Nigra.

 

Targeted devices

  • Kria KV260
  • Pynq Z2
  • MYIR FZ3
  • MYIR FZ5

 

Key capabilities

  • Training: Integration with Keras and Vitis AI Quantizer to prepare U-Net models for low-precision FPGA inference. The training flow is highly configurable, allowing for quick changes in the neural network and training parameters.

  • Vitis AI Flow: Full implementation of the Vitis AI compilation chain, transforming high-level models into hardware-optimized .xmodel files.

  • DPU Acceleration: Leveraging the Xilinx DPU for high-throughput, low-latency image segmentation.

  • Webapp: Includes a ready-to-use application for real-time inference and demonstration on the Kria KV260.

    • The application can also be run on the Pynq Z2, MYIR FZ3, and FZ5 boards.

  • Performance Benchmarking: Tools to evaluate inference accuracy and latency directly on the target hardware.

  • Software/Hardware Comparison: Both models are implemented in software flow and hardware-accelerated flow for direct comparison, and are available for use in the included web application. 
  • Ellipse regresor: The second part of the AI flow is implemented as a UNET model and a small ellipse regresor. Which will be used for further study of the echogenicity,

Features

  • Target Hardware: Specifically optimized for Xilinx Kria KV260, Pynq Z2, MYIR FZ3 and FZ5

  • End-to-End Pipeline: Covers everything from data preparation and training to quantization, compilation, and final deployment.

  • Hardware-Specific Scripts: Includes automated shell scripts for the entire Vitis AI flow (quantization, evaluation, and compilation).

  • Webapp: A complete web-based UI for interacting with the deployed model on the board.

What's included

  • Vitis AI Source: Comprehensive scripts (vitis_ai/) for the DPU compilation flow (scripts 0_... through 5_...).

  • Training Logic: Quantized U-Net implementation and training scripts compatible with FPGA deployment requirements.

  • Deployment Tools: The webapp source code and helper scripts for board-side execution.

Minimum requirements

  • OS: Desktop Linux (Ubuntu 18.04/20.04 recommended for Vitis AI tools).

  • Runtime: Python 3.7+; Vitis AI Docker environment or local toolchain (Quantizer, Compiler).

  • Hardware: * Development: Workstation with Xilinx Vitis AI toolchain installed.

    • Target: Xilinx Kria KV260, Pynq Z2, MYIR FZ3, MYIR FZ5 with DPU-based image.

  • Data: Raw images for segmentation

 

How to cite

Cite the Zenodo DOI for this version.

Authors: Denis Kurka, Petr Čermák

Files

denisuskurka/UNET_ACCEL-zenodo3.zip

Files (48.6 MB)

Name Size Download all
md5:2922f8f9f88affbc147006a07cc9ee3f
48.6 MB Preview Download

Additional details

Related works

Is supplement to
Software: https://github.com/denisuskurka/UNET_ACCEL/tree/zenodo3 (URL)

Funding

European Union
The project National Institute for Neurological Research, Programme EXCELES, Next Generation EU LX22NPO510x

Software

Repository URL
https://github.com/denisuskurka/UNET_ACCEL
Programming language
Python , C , C++ , VHDL
Development Status
Active

References

  • TensorFlow: Abadi, M., et al. (2015). TensorFlow: Large-scale machine learning on heterogeneous systems. https://www.tensorflow.org
  • Keras: Chollet, F., et al. (2015). Keras: The Python Deep Learning API. https://keras.io
  • U-Net (Original Paper): Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. arXiv preprint arXiv:1505.04597. https://doi.org/10.48550/arXiv.1505.04597
  • FINN (Original Paper): Umuroglu, Y., et al. (2017). FINN: A Framework for Fast, Scalable Binarized Neural Network Inference. Proceedings of the 2017 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA '17), 65–74. https://doi.org/10.1145/3020078.3021744
  • FINN-R (Journal Paper): Blott, M., et al. (2018). FINN-R: An end-to-end deep-learning framework for fast exploration of quantized neural networks. ACM Transactions on Reconfigurable Technology and Systems (TRETS), 11(3), 1–23. https://doi.org/10.1145/3242897