Back to Search
Start Over
Accelerating Plane Wave Imaging through Deep Learning-Based Reconstruction: An Experimental Study
- Source :
- 2017 IEEE International Ultrasonic Symposium (IUS), 2017 IEEE International Ultrasonic Symposium (IUS), Sep 2017, Washington, United States, 2017 IEEE International Ultrasonic Symposium (IUS), Sep 2017, Washington, United States. 2017, HAL
- Publication Year :
- 2017
- Publisher :
- HAL CCSD, 2017.
-
Abstract
- International audience; Background, Motivation and ObjectiveObtaining ultrafast images using steered plane wave (PW) imaging remains a challenge due to the trade-off between image quality and frame rate. PW imaging indeed relies on compounding in order to preserve a good image quality, usually using multiple successive emissions, which in turn yields a decrease of the frame rate. As opposed to this classicalapproach, we propose a new strategy to reduce the number of emitted PWs. This is done using a deep learning technique, i.e. by training a convolutional neural network (CNN) to reconstruct high quality images using a small number of PW emissions (typically two).Statement of Contribution/MethodsThe training data was generated as follows. A standard linear array probe (ATL L7-4 38 mm, 128 elements, bandwidth 4-7 MHz, transmitted frequency 5.208 MHz) was interfaced with a Verasonics system. 31 PWs were emitted (angle spanning ±15° in one degree steps). For each PW emission, the received raw-data was then processed using the f-k migration method [Garcia et al. IEEE UFFC13] to produce the reconstructed RF image. These 31 images were then averaged to get the final compounded image used as a reference in the training process, while only two images corresponding to angles -10° and +10° were given as input to the CNN.A total of 2000 reference images were acquired, 1000 images from in-vivo tissues (carotid and thyroid), 500 images from a Gammex phantom (410 SCG), and 500 images from a water-based polyamide mixture with air bubbles. 1800 images were used for training, 100 for validation, and 100 for testing. The CNN employed in this study is a fully-convolutionalneural network, with 4 hidden layers and maxout activation units. Training is done via stochastic gradient descent and the minimized loss function is the mean absolute error (MAE) between the CNN-reconstructed and the reference RF image, plus a L2 regularization term.Results/DiscussionOur approach was evaluated by comparing the CNN-based reconstruction of the test images with the conventional compounding method (see Fig. 1). Preliminary results indicate that CNN-based reconstruction is a very promising approach, as the reconstructed images are visually very close to those obtained by standard compounding with 31 PWs, while implying the emission of only 2 PWs.
- Subjects :
- [SPI.ACOU]Engineering Sciences [physics]/Acoustics [physics.class-ph]
[ SPI.ACOU ] Engineering Sciences [physics]/Acoustics [physics.class-ph]
[SDV.IB.IMA]Life Sciences [q-bio]/Bioengineering/Imaging
[ SPI.SIGNAL ] Engineering Sciences [physics]/Signal and Image processing
[SPI.SIGNAL]Engineering Sciences [physics]/Signal and Image processing
[ SDV.IB.IMA ] Life Sciences [q-bio]/Bioengineering/Imaging
Subjects
Details
- Language :
- English
- Database :
- OpenAIRE
- Journal :
- 2017 IEEE International Ultrasonic Symposium (IUS), 2017 IEEE International Ultrasonic Symposium (IUS), Sep 2017, Washington, United States, 2017 IEEE International Ultrasonic Symposium (IUS), Sep 2017, Washington, United States. 2017, HAL
- Accession number :
- edsair.dedup.wf.001..b8c635c4a3e6971fdfabef11dcdc00d2