GPU-Enabled Ultrasound Imaging Real-Time, Fully-Flexible Data Processing Dr. Christoph Hennersperger Research Manager, Technical University of Munich Research Fellow, Trinity College Dublin CTO, OneProjects
Manual Navigation Complex Diagnostics Reliability on Operator GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 2
Flexible Acquisition Guidance by/for Expert Intelligent Imaging GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 3
Real-Time, Fully-Flexible Data Acquisition Data Processing Toward Improved Diagnostics Dr. Christoph Hennersperger Research Manager, Technical University of Munich Research Fellow, Trinity College Dublin CTO, OneProjects
Brief Background on Ultrasound Imaging Workflow Beamforming Processing of received data Postprocessing for visualization Subsequent filters pipeline Electronic delays for focusing Radiofrequency data (RF) (e.g. speckle reduction) scanline signals Applicable in transmit and Reduction of dynamic range receive Envelope detection (Hilbert transform) (Log-compression) Transformation of scanlines to Subsampling (decimation) image (Scan-conversion) Carrier signal Scanline B-mode image Shape signal Scanline Scanline RF data Pulse shape Delay & Focus Scanline envelope Active elements GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 5
The Need for a Software Defined Ultrasound Framework Most ultrasound systems have limited flexibility Implementation of major processing on DSPs, FPGAs or ASICs Change of specific points require significant changes Most ultrasound systems are closed systems Access to images only through PACS Proprietary access and interfaces Not usable for fast prototyping or research GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 6
Software Defined Ultrasound Platform for Real-time Applications Mission: Provide framework to allow covering research aspects from low-level US to high-level applications Key Design Properties Data and module-driven approach Fully software-defined platform (with GPU) Fast prototyping Transparent storage Fully flexible design Fully real-time GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 7
General Ultrasound Processing Layout Receive Beamforming Transmit Beamforming Transmit & Receive (Delay and Sum (Aperture, Delays) with Apodization) Envelope detection incl. Log-Compression Scan-Conversion Frequency Compounding GPU accelerated Controllable via GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 8
Beamforming on GPU Parallelization for individual scanlines and samples Aperture defines input data (number of channels) + Delay and sum beamformer in SUPRA Blocks operate on receive scanlines τ 1 τ 2 τ 3 τ 4 τ 5 τ 6 τ 7 τ 8 Blocks process individual rx scanlines Shared memory over local aperture (memory access) Thread operate on receive samples Individual threads process samples of scanlines Local thread performs DAS over aperture (x,y) and depth (z) using shared memory GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 9
Imaging with SUPRA Direct Configuration Fast Acquisitions Storage of Full Data GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 10
Qualitative Evaluation to Proprietary Scanline Imaging Cephasonics SUPRA Muscle Fibers Point Phantom Carotid Longitudinal Carotid Transverse GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 11
From Scanline to Planewave Imaging Scanline Planewave Planewave 64 scanlines 64 Angles 10 Angles Max 300 Hz 300 Hz 1925 Hz GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 12
Real-time Capabilities of Framework Results for NVIDIA Jetson TX2, mobile GTX and GTX 1080 GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 13
Towards Improving Diagnostic Outcomes with US GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 14
Ultrasound - Unique Abilities and Challenges Data interpretation Image Graph (network) Continuous signals GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 15
Ultrasound - Unique Abilities and Challenges Data interpretation Image Graph (network) Continuous signals Understanding of physics Signals from reflection Acoustic tissue properties GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 16
Ultrasound - Unique Abilities and Challenges Data interpretation Image Graph (network) Continuous signals Understanding of physics Signals from reflection Acoustic tissue properties Challenges and artefacts Shadowing and enhancement Nonlinearity of tissue propagation Interference of waves GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 17
Modeling Ultrasound as Arbitrarily Sampled Data Overlapping US-slices • Resampling to regular grid • GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 19
Modeling Ultrasound as Arbitrarily Sampled Data Overlapping US-slices • Resampling to regular grid • Process on regular graph • Loss of information regarding acquisition! GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 20
Modeling Ultrasound as Arbitrarily Sampled Data Better: Define graph on original samples Graph nodes represent US samples • Graph edges represent spatial • structure Construct edges in “local” coordinates GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 21
Modeling Ultrasound as Arbitrarily Sampled Data Better: Define graph on original samples Graph nodes represent US samples • Graph edges represent spatial • structure Construct edges in “local” coordinates 1) Zu Berge, C. S., Declara, D., Hennersperger, C., Baust, M., & Navab, N. (2015, October). Real-time uncertainty visualization for B-mode ultrasound. IEEE SciVis 2015. 2) Hennersperger, C., Mateus, D., Baust, M., & Navab, N. (2014, September). A quadratic energy minimization framework for signal loss estimation from arbitrarily sampled ultrasound data, MICCAI 2015 3) Virga, S., Zettinig, O., Esposito, M., Pfister, K., Frisch, … & Hennersperger, C. (2016, October). Automatic force-compliant robotic ultrasound screening of abdominal aortic aneurysms, IEEE IROS 2016 GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 23
Modeling Ultrasound as Arbitrarily Sampled Data Better: Define graph on original samples Graph nodes represent US samples • Graph edges represent spatial • structure Construct edges in “local” coordinates 1) Virga, S., Göbl, R., Baust, M., Navab, N., & Hennersperger, C. (2018). Use the force: deformation correction in robotic 3D ultrasound. International journal of computer assisted radiology and surgery, 1-9. GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 24
Connecting the Dots Improving Diagnostic Outcomes Domain specific knowledge to improve diagnostic benefit Data science approaches with need for data (end to end) SUPRA as enabling technology Fully open, flexible, and real-time • Tool for rapid demonstration and R&D • GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 25
Next: Try it Yourself! Christoph Hennersperger Rüdiger Göbl Nassir Navab Even without an ultrasound machine: https://github.com/IFL-CAMP/supra This project received funding from the European Union’s H2020 research and innovation programme (No 688279 GPU-Enabled Ultrasound Imaging | Christoph Hennersperger 1/27/2016 | Slide 26
Recommend
More recommend
