Deep Learning for Ultrasound Beamforming in Flexible Array Transducer

Area of Science:

• Medical Imaging
• Ultrasound Technology
• Artificial Intelligence in Medicine

Background:

• Flexible array transducers offer advantages for image-guided radiotherapy tumor tracking.
• Conventional delay-and-sum (DAS) beamforming struggles with flexible transducer geometry, causing image distortion and defocusing.

Purpose of the Study:

• To develop a novel deep learning approach to overcome limitations of DAS beamforming with unknown flexible transducer geometries.
• To reconstruct high-quality, undistorted B-mode ultrasound images directly from radio-frequency (RF) data.

Main Methods:

• An end-to-end deep neural network (DNN) approach was designed to learn optimal time delays for RF data channels.
• The DNN beamformer was compared against the conventional DAS beamformer using simulations and flexible array transducer scan data.

Main Results:

• The DNN approach significantly reduced the full-width-at-half-maximum (FWHM) of point scatterers (1.80 mm in simulation, 1.31 mm in scans).
• Contrast-to-noise ratio (CNR) for anechoic cysts improved by 0.79 dB (simulation) and 1.69 dB (phantom scans).
• Image distortion was reduced, with cyst aspect ratios closer to 1, indicating improved accuracy.

Conclusions:

• The proposed DNN approach effectively addresses distortion and improves lateral resolution and contrast in B-mode ultrasound images from flexible array transducers.
• This method holds promise for enhancing image quality in applications like image-guided radiotherapy.