Personal tools

Ultrasound image data recycler

From iis-projects

Jump to: navigation, search
USDataRecycler image.png

Short Description

In clinics and in biomechanical research, it is common to take ultrasound images of muscles and tendons. Therefore, huge digital data sets exist. Since medical ultrasound equipment used is proprietary and closed, it is not possible to access the raw data of these images. In novel robotics hardware with wearable ultrasonic sensors, e.g., for prosthesis control, networks are used for sensor control at the edge. Training these models would benefit from raw data, but it is scarce. To overcome this dilemma, this project aims to develop an image-to-raw data converter. A mixed deterministic/data-driven model is implemented in C++ and Python. The focus is on developing a "physically informed" model that includes an ultrasound simulator (deterministic), a CNN U-net model (data-driven), and a joint loss function to convert images to raw data.

Goal & Tasks

  • Setup of the ultrasound simulator using the C++ implementation of k-wave.
  • Implementing a data generator to produce random density maps and linking the generator to the ultrasound simulator.
  • Adapt a Convolutional Neural Network (U-Net) to segment ultrasound images into density maps.
  • Combine both models (simulator and CNN) and penalize the CNN output with the original density map input of the simulator.



  • Machine Learning
  • Python
  • C++

Status: Available

Supervision: Christoph Leitner, Martin Danelljan


20% Literature Study
60% model development
20% training and testing


Luca Benini

Practical Details

↑ top