A Novel End-To-End Deep Learning Based Beamformer for Flexible Array Ultrasound Transducer
Presentations
WE-F-TRACK 3-3 (Wednesday, 7/28/2021) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Purpose: The flexible array transducer is a promising tool for image-guided tumor tracking in radiotherapy. Comparing with the traditional ultrasound transducer, it is user-independent and can avoid contact pressure that may cause an anatomical change. However, due to its flexible geometry, the conventional delay-and-sum (DAS) beamformer may apply incorrect time delay to the radio-frequency (RF) data and produce B-mode images with low quality and considerable distortion. Here, we propose a novel end-to-end deep learning approach that may alternate the conventional DAS beamformer when the transducer geometry is unknown.
Methods: Our deep neural networks (DNNs) are designed based on the popular U-Net, Pix2Pix GAN, and Cycle GAN structures. They are trained with Field II simulation and real scanning RF data from transducers with different geometries and the corresponding properly beamformed images. The DNNs are expected to learn the proper time delays for each channel and generate the undistorted high-quality B-mode images directly from RF data. They are then tested with RF data from a transducer with random unknown geometry, and the results are evaluated using full-width-at-half-maximum (FWHM), contrast-to-noise ratio (CNR), and aspect ratio of the round-shaped cyst.
Results: We compare the DNN results to the standard DAS results using different data. For point targets, the average FWHM of DNN results is 2.1 mm lower than that of DAS results. For anechoic cyst targets, the DNN results improve the CNR by 0.79 dB and aspect ratio by 0.23. For abdominal phantom and in vivo scans, the DNN approach improves the CNR by 1.80 dB and 2.18 dB, respectively. The U-Net structure has the overall best performance.
Conclusion: The evaluations show that when the transducer geometry is unknown, the proposed DNNs can effectively reduce the distortion and improve the lateral resolution and contrast of the B-mode images compare with the conventional DAS beamformer.
Funding Support, Disclosures, and Conflict of Interest: Research Supported by National Cancer Institute R37CA229417
Handouts
Keywords
Ultrasonics, Transducers, Reconstruction
Taxonomy
IM- Ultrasound : Machine learning, computer vision
Contact Email