Home Hardware dependencies of GPU-accelerated beamformer performances for microwave breast cancer detection
Article Open Access

Hardware dependencies of GPU-accelerated beamformer performances for microwave breast cancer detection

  • Christoph J. Salomon EMAIL logo , Sebastian Ley and Marko Helbig
Published/Copyright: September 30, 2016

Abstract

UWB microwave imaging has proven to be a promising technique for early-stage breast cancer detection. The extensive image reconstruction time can be accelerated by parallelizing the execution of the underlying beamforming algorithms. However, the efficiency of the parallelization will most likely depend on the grade of parallelism of the imaging algorithm and of the utilized hardware. This paper investigates the dependencies of two different beamforming algorithms on multiple hardware specification of several graphics boards. The parallel implementation is realized by using NVIDIA’s CUDA. Three conclusions are drawn about the behavior of the parallel implementation and how to efficiently use the accessible hardware.

1 Introduction

Ultra-wideband (UWB) microwave imaging is a promising method for early-stage breast cancer detection, which has been successfully tested by numerous research groups [1], [2], [3], [4]. It is based on the differing dielectric properties of tumors and surrounding tissue, originating from their different water content.

For imaging purposes, UWB electromagnetic stimuli are sent into the material under test. Beamforming algorithms are then applied to calculate the reflected energy for numerous focal points inside the area of interest. The tumor detection requires an adequate resolution of the scanning grid, to ensure that all necessary dielectric transitions inside the breast are detected. Depending on the desired image resolution, the number of channels and the complexity of the applied beamformer, respectively, the time for a 3D-Image reconstruction will elevate rapidly.

One approach to accelerate these computations is to shift the imaging process from serial execution on a Central Processing Unit (CPU, also referred to as host) to massively parallel execution on a Graphics Processing Unit (GPU, also referred to as device). This paradigm is known as General Purpose Programming of Graphics Processing Units (GPGPU). Utilizing the many-core architecture of GPUs, several stages of the imaging process can be computed simultaneously by different computational cores, resulting in a massive reduction of the image reconstruction time [5]. However, various types of beamformers are available for breast cancer detection, all differing in characteristics and complexity as described in [1]. To make efficient use of GPGPU, the accelerating hardware should match the achievable parallelism of the applied algorithm.

This paper presents a parallel solution for a Delay-and-sum (DAS) beamformer and a Delay-multiply-and-sum (DMAS) beamformer, respectively, and the acceleration achieved for different types of graphics boards. The boards vary in their GPU’s processor architecture (Maxwell and Kepler), the amount of their GPU’s processor cores (CUDA-Cores), their clock rate and their memory bandwidth, respectively. The runtime of the parallel execution is presented as a function of pixel quantity and benchmarked against a fully serial implementation computed by a general purpose CPU. To utilize the GPU, NVIDIA’s compute unified device architecture (CUDA) provides the necessary programming interface. In the following sections we describe the applied beamforming algorithms with respect to their potential parallelism, our way of implementation and the specifications that may have influence on the overall performance. We present the achieved acceleration on different hardware architectures and draw a conclusion about factors that have to be considered for an efficient use of parallel hardware when accelerating beamforming algorithms for medical UWB imaging.

2 Imaging algorithm

The DAS algorithm is the most basic implementation of a beamformer. Based on [4] it can mathematically be described as

(1)I(r0)=[n=1NSn(τn(r0))]2.

I(r0) is the backscattered energy of the focal point r0, with r0 = [x0, y0, z0]. Sn(t) represents the clutter subtracted signal of channel n, τn the corresponding signal runtime along the path Txn-r0-Rxn (Txn: transmitting antenna of channel n; Rxn: receiving antenna of channel n) and N the total number of channels. To calculate the reflected energy of an arbitrary point inside the area of interest, the Euclidean distance between the antennas and the coordinates of the focal point has to be calculated. Considering signal propagation time and sampling rate, the representing sample of the respective channel for this point is identified. To increase the signal-to-noise ratio (SNR), these steps are repeated for numerous channels and the results are summed up. Iterating through a matrix of focal points and color coding the results creates a visually interpretable image of reflected signal energy inside the breast.

To further increase the SNR, Lim et al. [6] introduced the DMAS algorithm which contains an additional pairwise multiplication of all channels. Deriving from their presented diagram the DMAS can mathematically be described as

(2)I(r0)=[n=1N1m=n+1NSn(τn(r0))Sm(τm(r0))]2.

By examining equations (1) and (2) it becomes obvious, that the reflected energy for all focal points can be computed independently. However, the calculation itself contains serial segments that cannot be parallelized. To sum all intermediate results, the same memory space has to be addressed at one point of the calculation for each channel, which limits the achievable parallelism of the overall calculation. The additional pairwise multiplication and addition inside the DMAS algorithm restricts this parallelism even more. The ratio of serial and parallel executable parts differs for both algorithms. Therefore, the efficiency of the parallel solution for the respective beamforming algorithm will most likely depend on the grade of parallelism of the utilized hardware.

3 Parallel implementation

The following section describes the specifications and settings of our CUDA implementation that may have influence on the overall performance of the parallel execution, including data transfer and hardware initialization.

Since host and device memory are physically separated, all required data has to be transferred to the device memory first, before it can be processed by the GPU. The backscattered signals and coordinates of antennas and focal points are copied to the device before the first image point is computed. Results are transferred back to the host memory only after all computations of the graphics card have been completed. Due to intelligent caching behavior, no memories of the GPU except for global memory (DRAM) are used. All data is kept in page-locked memory on the host before being copied. Once the data is available to the device, the parallel computation is started.

To utilize graphics boards for general-purpose computations, CUDA maps the physical computation cores of the GPU to a three-dimensional grid of threads. Threads are then again organized in blocks. The number of threads per block has direct impact on the processor occupancy and the achieved acceleration. A block-size of 8 × 8 × 1 has proven to be most efficient in average for multiple graphics cards and is used for all our computations. As the number of computed image points varies, the grid-size is adapted in a way that every thread computes one single image point. Both beamformers are implemented as a function that is executed by every thread, the so-called kernel. The kernel processes all necessary steps described in (1) and (2), respectively.

4 Accelerated püerformance

To illustrate the dependencies of reconstruction time as a function of pixel quantity, different amounts of image layers are reconstructed during a single program execution. Although the scanning matrix can be chosen arbitrarily, for the purpose of comparison every reconstructed image layer lies perpendicular to the axis of rotational symmetry of the breast phantom and consists of 10201 pixels. This matches a resolution of 1 mm per plane for an image dimension of 10 cm × 10 cm. Furthermore, every image point is computed applying 4096 back scattered signals resulting from 128 channels as described in [2].

A benchmark is set by first executing a complete serial implementation of both beamformers on a general purpose CPU of type Intel Core i5 330. The results of this execution are presented in Figure 1. The image reconstruction is executed on a single computation core at 3.33 GHz clock rate. Recordings of runtime are implemented in the serial code using time stamp functions of the C standard-libraries.

Figure 1 Execution time of serial DAS (top) and DMAS (bottom), respectively on a single computation core of a CPU type Intel Core i5 330 at 3.33 GHz. Runtime is presented as a function of reconstructed image layers. Each layer consists of 10201 pixels, corresponding to 1 mm image resolution for an image dimension of 10 cm × 10 cm.
Figure 1

Execution time of serial DAS (top) and DMAS (bottom), respectively on a single computation core of a CPU type Intel Core i5 330 at 3.33 GHz. Runtime is presented as a function of reconstructed image layers. Each layer consists of 10201 pixels, corresponding to 1 mm image resolution for an image dimension of 10 cm × 10 cm.

Since the kernel runs asynchronous to the host code, casual timestamps are not applicable to determine the parallel execution time. For that purpose CUDA events were used. The achieved acceleration is illustrated in Figure 2 for different graphics boards. Acceleration refers to the results presented in Figure 1 and is the ratio of serial execution time and parallel execution time, including the time necessary for copying data between host and device. During the execution we tried to minimize the number of running processes and interrupts. The exact hardware specifications of all utilized graphics boards can be seen in Table 1.

Figure 2 Achieved acceleration of parallel beamformer execution. Top (A): DAS Beamformer. Bottom (B): DMAS Beamformer. Acceleration refers to the results from Figure 1 and is the ratio of serial execution time and parallel execution time. Triangles represent GPUs with Kepler architecture; Dots refer to GPUs with Maxwell architecture.
Figure 2

Achieved acceleration of parallel beamformer execution. Top (A): DAS Beamformer. Bottom (B): DMAS Beamformer. Acceleration refers to the results from Figure 1 and is the ratio of serial execution time and parallel execution time. Triangles represent GPUs with Kepler architecture; Dots refer to GPUs with Maxwell architecture.

Table 1

Hardware specifications of utilized graphics boards.

GTX 750 TiGTX 760GTX 960
CUDA-Cores64011521024
Clock (MHz)10249801127
Memory-bandwidth (GB/s)86.4192112
GPU architectureMaxwellKeplerMaxwell
DRAM (MB)204820482048

Figure 3 illustrates a reconstructed image layer using a DAS and a DMAS algorithm, respectively. A tumor imitation is placed inside the breast phantom as described in [2].

Figure 3 Reconstructed image layer of breast phantom with a tumor imitate of 15 mm diameter. Resolution is 1 mm. Top: DAS-algorithm applied. Bottom: DMAS-algorithm applied. Linear scale of energy values.
Figure 3

Reconstructed image layer of breast phantom with a tumor imitate of 15 mm diameter. Resolution is 1 mm. Top: DAS-algorithm applied. Bottom: DMAS-algorithm applied. Linear scale of energy values.

5 Conclusion

Evaluating the results, we can draw three conclusions about how to use the presented hardware efficiently for GPU-accelerated microwave breast imaging.

First: The computation shifted to the GPU has to be as extensive as possible. Figure 2(A) shows, that with increasing problem size more acceleration is achieved until a certain point of saturation. For an effective use the hardware has to operate at full capacity. Due to CUDA’s intelligent thread-management this capacity cap is not the amount of computation cores but has to be determined empirically.

Second: The implementation performance of a highly parallel algorithm is mostly dependent on the number of processing cores available on the accelerating hardware. Apparently, the highest acceleration is achieved with the GTX 760, which offers the lowest clock rate but the highest number of CUDA-Cores. Therefore, clock rate is the least significant factor for an effective choice of hardware. However, the GTX 760 also offers the highest memory bandwidth. Examining Figure 2(A) it appears that with higher problem size, the acceleration achieved with the GTX 970 approaches the GTX 760’s performance. That is, even the GTX 760 features a much higher memory bandwidth. The gap between the amounts of computational cores however, is a lot smaller. We conclude, that the number of computational cores has a higher impact on the performance, than the memory bandwidth. When executing a less parallel problem, the decisive component changes, which leads to our next conclusion.

Third: An accelerated problem that still contains a lot of serial computations after the parallelisation is more efficiently executed by NVIDIA GPUs with Kepler architecture than by GPUs with Maxwell architecture. When applying the DMAS algorithm, again the GTX 760 scores the best results. GTX 960 and GTX 750 show nearly equal results, even though their hardware specifications differ a lot. However, both GPUs have the same hardware architecture. The main difference between the graphics boards disregarding number of computation cores, memory bandwidth and clock is the GPU architecture.

All presented imaging algorithms imply independency of the computed image points among each other. However, some algorithms include certain dependencies among neighboring focal points. Future investigations have to focus on efficient ways of accelerating these methods, where our approach of parallelisation is not applicable. Additionally, a broader study of graphics cards will probably lead to a more general conclusion about efficient use of accelerating hardware.

Author’s Statement

Conflict of interest: Authors state no conflict of interest. Material and Methods: Informed consent: Informed consent has been obtained from all individuals included in this study. Ethical approval: The research related to human use complies with all the relevant national regulations, institutional policies and was performed in accordance with the tenets of the Helsinki Declaration, and has been approved by the authors’ institutional review board or equivalent committee.

References

[1] Byrne D, O’Halloran M, Glavin M, Jones E. Data independent radar beamforming algorithms for breast cancer detection. Prog Electromagn Res. 2010;107:331–48.10.2528/PIER10061001Search in Google Scholar

[2] Helbig M, Dahlke K, Hilger I, Kmec M, Sachs J. Design and test of an imaging system for UWB breast cancer detection. Frequenz. 2012;66:387–94.10.1515/freq-2012-0103Search in Google Scholar

[3] Klemm M, Leendertz JA, Gibbins D, Craddock IJ, Preece A, Benjamin R. Microwave radar-based differential breast cancer imaging: imaging in homogeneous breast phantoms and low contrast scenarios. IEEE Trans Antennas Propag. 2010;58:2337–44.10.1109/TAP.2010.2048860Search in Google Scholar

[4] Li X, Davis SK, Hagness SC, Van der Weide DW, Van Veen BD. Microwave imaging via space-time beamforming: experimental investigation of tumor detection in multilayer breast phantoms. IEEE Trans Microw Theory Tech. 2004;52:1856–65.10.1109/TMTT.2004.832686Search in Google Scholar

[5] Elahi MA, Shahzad A, Glavin M, Jones E, O’Halloran M. GPU accelerated confocal microwave imaging algorithms for breast cancer detection. The 9th European IEEE Conference on Antennas and Propagation (EuCAP) 2015:1–2.Search in Google Scholar

[6] Lim HB, Nhung NTT, Li EP, Thang ND. Confocal microwave imaging for breast cancer detection: delay-multiply-and-sum image reconstruction algorithm. IEEE Trans Biomed Eng. 2008;55:1697–704.10.1109/TBME.2008.919716Search in Google Scholar

Published Online: 2016-9-30
Published in Print: 2016-9-1

©2016 Christoph J. Salomon et al., licensee De Gruyter.

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Articles in the same Issue

  1. Synthesis and characterization of PIL/pNIPAAm hybrid hydrogels
  2. Novel blood protein based scaffolds for cardiovascular tissue engineering
  3. Cell adhesion and viability of human endothelial cells on electrospun polymer scaffolds
  4. Effects of heat treatment and welding process on superelastic behaviour and microstructure of micro electron beam welded NiTi
  5. Long-term stable modifications of silicone elastomer for improved hemocompatibility
  6. The effect of thermal treatment on the mechanical properties of PLLA tubular specimens
  7. Biocompatible wear-resistant thick ceramic coating
  8. Protection of active implant electronics with organosilicon open air plasma coating for plastic overmolding
  9. Examination of dielectric strength of thin Parylene C films under various conditions
  10. Open air plasma deposited antimicrobial SiOx/TiOx composite films for biomedical applications
  11. Systemic analysis about residual chloroform in PLLA films
  12. A macrophage model of osseointegration
  13. Towards in silico prognosis using big data
  14. Technical concept and evaluation of a novel shoulder simulator with adaptive muscle force generation and free motion
  15. Usability evaluation of a locomotor therapy device considering different strategies
  16. Hypoxia-on-a-chip
  17. Integration of a semi-automatic in-vitro RFA procedure into an experimental setup
  18. Fabrication of MEMS-based 3D-μECoG-MEAs
  19. High speed digital interfacing for a neural data acquisition system
  20. Bionic forceps for the handling of sensitive tissue
  21. Experimental studies on 3D printing of barium titanate ceramics for medical applications
  22. Patient specific root-analogue dental implants – additive manufacturing and finite element analysis
  23. 3D printing – a key technology for tailored biomedical cell culture lab ware
  24. 3D printing of hydrogels in a temperature controlled environment with high spatial resolution
  25. Biocompatibility of photopolymers for additive manufacturing
  26. Biochemical piezoresistive sensors based on pH- and glucose-sensitive hydrogels for medical applications
  27. Novel wireless measurement system of pressure dedicated to in vivo studies
  28. Portable auricular device for real-time swallow and chew detection
  29. Detection of miRNA using a surface plasmon resonance biosensor and antibody amplification
  30. Simulation and evaluation of stimulation scenarios for targeted vestibular nerve excitation
  31. Deep brain stimulation: increasing efficiency by alternative waveforms
  32. Prediction of immediately occurring microsleep events from brain electric signals
  33. Determining cardiac vagal threshold from short term heart rate complexity
  34. Classification of cardiac excitation patterns during atrial fibrillation
  35. An algorithm to automatically determine the cycle length coverage to identify rotational activity during atrial fibrillation – a simulation study
  36. Deriving respiration from high resolution 12-channel-ECG during cycling exercise
  37. Reducing of gradient induced artifacts on the ECG signal during MRI examinations using Wilcoxon filter
  38. Automatic detection and mapping of double potentials in intracardiac electrograms
  39. Modeling the pelvic region for non-invasive pelvic intraoperative neuromonitoring
  40. Postprocessing algorithm for automated analysis of pelvic intraoperative neuromonitoring signals
  41. Best practice: surgeon driven application in pelvic operations
  42. Vasomotor assessment by camera-based photoplethysmography
  43. Classification of morphologic changes in photoplethysmographic waveforms
  44. Novel computation of pulse transit time from multi-channel PPG signals by wavelet transform
  45. Efficient design of FIR filter based low-pass differentiators for biomedical signal processing
  46. Nonlinear causal influences assessed by mutual compression entropy
  47. Comparative study of methods for solving the correspondence problem in EMD applications
  48. fNIRS for future use in auditory diagnostics
  49. Semi-automated detection of fractional shortening in zebrafish embryo heart videos
  50. Blood pressure measurement on the cheek
  51. Derivation of the respiratory rate from directly and indirectly measured respiratory signals using autocorrelation
  52. Left cardiac atrioventricular delay and inter-ventricular delay in cardiac resynchronization therapy responder and non-responder
  53. An automatic systolic peak detector of blood pressure waveforms using 4th order cumulants
  54. Real-time QRS detection using integrated variance for ECG gated cardiac MRI
  55. Preprocessing of unipolar signals acquired by a novel intracardiac mapping system
  56. In-vitro experiments to characterize ventricular electromechanics
  57. Continuous non-invasive monitoring of blood pressure in the operating room: a cuffless optical technology at the fingertip
  58. Application of microwave sensor technology in cardiovascular disease for plaque detection
  59. Artificial blood circulatory and special Ultrasound Doppler probes for detecting and sizing gaseous embolism
  60. Detection of microsleep events in a car driving simulation study using electrocardiographic features
  61. A method to determine the kink resistance of stents and stent delivery systems according to international standards
  62. Comparison of stented bifurcation and straight vessel 3D-simulation with a prior simulated velocity profile inlet
  63. Transient Euler-Lagrange/DEM simulation of stent thrombosis
  64. Automated control of the laser welding process of heart valve scaffolds
  65. Automation of a test bench for accessing the bendability of electrospun vascular grafts
  66. Influence of storage conditions on the release of growth factors in platelet-rich blood derivatives
  67. Cryopreservation of cells using defined serum-free cryoprotective agents
  68. New bioreactor vessel for tissue engineering of human nasal septal chondrocytes
  69. Determination of the membrane hydraulic permeability of MSCs
  70. Climate retainment in carbon dioxide incubators
  71. Multiple factors influencing OR ventilation system effectiveness
  72. Evaluation of an app-based stress protocol
  73. Medication process in Styrian hospitals
  74. Control tower to surgical theater
  75. Development of a skull phantom for the assessment of implant X-ray visibility
  76. Surgical navigation with QR codes
  77. Investigation of the pressure gradient of embolic protection devices
  78. Computer assistance in femoral derotation osteotomy: a bottom-up approach
  79. Automatic depth scanning system for 3D infrared thermography
  80. A service for monitoring the quality of intraoperative cone beam CT images
  81. Resectoscope with an easy to use twist mechanism for improved handling
  82. In vitro simulation of distribution processes following intramuscular injection
  83. Adjusting inkjet printhead parameters to deposit drugs into micro-sized reservoirs
  84. A flexible standalone system with integrated sensor feedback for multi-pad electrode FES of the hand
  85. Smart control for functional electrical stimulation with optimal pulse intensity
  86. Tactile display on the remaining hand for unilateral hand amputees
  87. Effects of sustained electrical stimulation on spasticity assessed by the pendulum test
  88. An improved tracking framework for ultrasound probe localization in image-guided radiosurgery
  89. Improvement of a subviral particle tracker by the use of a LAP-Kalman-algorithm
  90. Learning discriminative classification models for grading anal intraepithelial neoplasia
  91. Regularization of EIT reconstruction based on multi-scales wavelet transforms
  92. Assessing MRI susceptibility artefact through an indicator of image distortion
  93. EyeGuidance – a computer controlled system to guide eye movements
  94. A framework for feedback-based segmentation of 3D image stacks
  95. Doppler optical coherence tomography as a promising tool for detecting fluid in the human middle ear
  96. 3D Local in vivo Environment (LivE) imaging for single cell protein analysis of bone tissue
  97. Inside-Out access strategy using new trans-vascular catheter approach
  98. US/MRI fusion with new optical tracking and marker approach for interventional procedures inside the MRI suite
  99. Impact of different registration methods in MEG source analysis
  100. 3D segmentation of thyroid ultrasound images using active contours
  101. Designing a compact MRI motion phantom
  102. Cerebral cortex classification by conditional random fields applied to intraoperative thermal imaging
  103. Classification of indirect immunofluorescence images using thresholded local binary count features
  104. Analysis of muscle fatigue conditions using time-frequency images and GLCM features
  105. Numerical evaluation of image parameters of ETR-1
  106. Fabrication of a compliant phantom of the human aortic arch for use in Particle Image Velocimetry (PIV) experimentation
  107. Effect of the number of electrodes on the reconstructed lung shape in electrical impedance tomography
  108. Hardware dependencies of GPU-accelerated beamformer performances for microwave breast cancer detection
  109. Computer assisted assessment of progressing osteoradionecrosis of the jaw for clinical diagnosis and treatment
  110. Evaluation of reconstruction parameters of electrical impedance tomography on aorta detection during saline bolus injection
  111. Evaluation of open-source software for the lung segmentation
  112. Automatic determination of lung features of CF patients in CT scans
  113. Image analysis of self-organized multicellular patterns
  114. Effect of key parameters on synthesis of superparamagnetic nanoparticles (SPIONs)
  115. Radiopacity assessment of neurovascular implants
  116. Development of a desiccant based dielectric for monitoring humidity conditions in miniaturized hermetic implantable packages
  117. Development of an artifact-free aneurysm clip
  118. Enhancing the regeneration of bone defects by alkalizing the peri-implant zone – an in vitro approach
  119. Rapid prototyping of replica knee implants for in vitro testing
  120. Protecting ultra- and hyperhydrophilic implant surfaces in dry state from loss of wettability
  121. Advanced wettability analysis of implant surfaces
  122. Patient-specific hip prostheses designed by surgeons
  123. Plasma treatment on novel carbon fiber reinforced PEEK cages to enhance bioactivity
  124. Wear of a total intervertebral disc prosthesis
  125. Digital health and digital biomarkers – enabling value chains on health data
  126. Usability in the lifecycle of medical software development
  127. Influence of different test gases in a non-destructive 100% quality control system for medical devices
  128. Device development guided by user satisfaction survey on auricular vagus nerve stimulation
  129. Empirical assessment of the time course of innovation in biomedical engineering: first results of a comparative approach
  130. Effect of left atrial hypertrophy on P-wave morphology in a computational model
  131. Simulation of intracardiac electrograms around acute ablation lesions
  132. Parametrization of activation based cardiac electrophysiology models using bidomain model simulations
  133. Assessment of nasal resistance using computational fluid dynamics
  134. Resistance in a non-linear autoregressive model of pulmonary mechanics
  135. Inspiratory and expiratory elastance in a non-linear autoregressive model of pulmonary mechanics
  136. Determination of regional lung function in cystic fibrosis using electrical impedance tomography
  137. Development of parietal bone surrogates for parietal graft lift training
  138. Numerical simulation of mechanically stimulated bone remodelling
  139. Conversion of engineering stresses to Cauchy stresses in tensile and compression tests of thermoplastic polymers
  140. Numerical examinations of simplified spondylodesis models concerning energy absorption in magnetic resonance imaging
  141. Principle study on the signal connection at transabdominal fetal pulse oximetry
  142. Influence of Siluron® insertion on model drug distribution in the simulated vitreous body
  143. Evaluating different approaches to identify a three parameter gas exchange model
  144. Effects of fibrosis on the extracellular potential based on 3D reconstructions from histological sections of heart tissue
  145. From imaging to hemodynamics – how reconstruction kernels influence the blood flow predictions in intracranial aneurysms
  146. Flow optimised design of a novel point-of-care diagnostic device for the detection of disease specific biomarkers
  147. Improved FPGA controlled artificial vascular system for plethysmographic measurements
  148. Minimally spaced electrode positions for multi-functional chest sensors: ECG and respiratory signal estimation
  149. Automated detection of alveolar arches for nasoalveolar molding in cleft lip and palate treatment
  150. Control scheme selection in human-machine- interfaces by analysis of activity signals
  151. Event-based sampling for reducing communication load in realtime human motion analysis by wireless inertial sensor networks
  152. Automatic pairing of inertial sensors to lower limb segments – a plug-and-play approach
  153. Contactless respiratory monitoring system for magnetic resonance imaging applications using a laser range sensor
  154. Interactive monitoring system for visual respiratory biofeedback
  155. Development of a low-cost senor based aid for visually impaired people
  156. Patient assistive system for the shoulder joint
  157. A passive beating heart setup for interventional cardiology training
Downloaded on 12.9.2025 from https://www.degruyterbrill.com/document/doi/10.1515/cdbme-2016-0111/html
Scroll to top button