Deep Learning in Medical Ultrasound Imaging

During the last decade, deep learning has had a profound impact on processing complex data and making associated decisions.It demonstrates that networks can be trained from a large number of examples and thereby enables us to solve extremely complicated problems which are nearly impossible to be handled by traditional logic programming or model-based approaches.

The accessibility of deep learning has also been considerably eased by comprehensive open source libraries available such as Google tensorflow, Pytorch, Caffe, etc.

Deep learning is applicable to ultrasound signals as well as to standard images.

This PhD project will address two topics in improving ultrasound imaging. The first is Super Resolution Imaging (SRI) through detecting and tracking micro bubbles. A contrast agent is injected into blood stream and the bubbles are tracked on the image. This leads to very high precision ultrasound images resulting in the higher resolution of 10 μm compared to the normal 0.5 mm resolution.

It is, however, difficult to get a reliable identification and isolation of the bubbles, and track them on images of complex vascular structures. To cope with these difficulties, deep learning networks can be trained for better and more stable isolation and tracking of bubbles based on either simulation results from Field II or real measurements. This will demand heavy computations, and DTU Elektro and DTU Compute have invested in a 3000 core Intel cluster and a GPU cluster based on the Volta NVIDIA GPU.

The second part of this project is related to ultrasound beamforming. Ultrasound data are acquired using 128 to 192 element arrays through transducer. All these signals are combined using delay-and-sum beamforming. It is a very simple, but computationally efficient method, which can be conducted in real-time.

This method, however, limits the attainable resolution and contrast, therefore, minimum variance schemes which are more complex have been developed. They can increase resolution from a wavelength to a tenth of a wavelength for isolated targets but fail to work well for many target cases typically found in medical ultrasound.

The idea is to replace the beamformer with a deep neural network so that we can obtain decent results even for many target cases. The network will be trained on a range of target examples from both Field II simulation results and measured signals using the experimental ultrasound scanner SARUS. It will then be investigated how many and how dense the targets can be, and the method will also be employed on in-vivo ultrasound signals acquired from SARUS.

This project demands knowledge form both medical ultrasound and deep learning, therefore, it is being conducted in a close collaboration between DTU Elektro and DTU Compute.