[home] [Personal Program] [Help]

---

Session: Vascular I
Session starts: Thursday 26 January, 14:30
Presentation starts: 15:30
Room: Room 559

---

Nathan Blanken (University of Twente)
Alina Kuliesh (Delft University of Technology)
Baptiste Heiles (Delft University of Technology)
Kartik Jain (University of Twente)
Hervé Delingette (INRIA)
Christoph Brune (University of Twente)
Michel Versluis (University of Twente)
David Maresca (Delft University of Technology)
Jelmer M. Wolterink (University of Twente)
Guillaume Lajoinie (University of Twente)

Abstract:
Ultrasound is a cost-effective, non-ionizing, portable imaging technique offering excellent temporal resolution. In blood flow imaging, the poor contrast can be addressed by administering contrast agents in the form of a microbubble suspension. Ultrasound vector flow imaging (echoPIV) has shown an accuracy on par with phase contrast MRI. However, diagnosis of vascular diseases would benefit substantially from further improved resolution, beyond the diffraction limit of ultrasound. Current ultrasound super-resolution methods require long acquisition times and are often based on microbubble localization from reconstructed grey-scale images. Here, we introduce a new super-resolution ultrasound technique based on the deconvolution of raw radio-frequency (RF) ultrasound data using deep learning. As a first step, we perform super-resolution before image reconstruction. Our strategy consists of: i) generating ultrasound RF data together with the ground-truth bubble locations with a custom, fast simulator and ii) training a neural network to detect bubbles within the simulated data. The simulator defines a random spatial microbubble distribution and simulates its response to a propagating plane wave. We use a dilated convolution neural network to recover the bubble locations within a single RF channel. In a future step, we will perform super-resolution and image reconstruction jointly, using end-to-end neural network training on raw ultrasound data. To this end, we have developed a state-of-the-art, physically realistic simulator that captures arterial flow, nonlinear acoustics, and nonlinear microbubble behaviour. Application of a trained neural network to single-element ultrasound data yields an order-of-magnitude gain in axial resolution in the final reconstructed images compared to standard B-mode images. Moreover, our method shows a substantial improvement in localization accuracy compared to other super-resolution methods, owing to accurate simulation of the microbubble response and access to the ground truth locations during training. Our results demonstrate that the application of a convolutional neural network to raw transducer element data is a promising path towards super-resolved imaging of high-density microbubble populations, allowing for real-time, high-resolution blood flow imaging and improved diagnosis of vascular diseases.