Speckle reduction of medical ultrasound images using deep learning with fully convolutional network

Abstract

Smoothing filters are frequently used for speckle reduction of medical ultrasound images. However, such filters may cause loss of the detailed structures of tissues in terms of image contrast. To improve image contrast in speckle reduction, we investigated a filter for medical ultrasound images using deep learning with a fully convolutional network, which was trained with pairs of input and target data generated by computer simulation. The proposed method achieved higher contrast-to-noise ratio and contrast values than the conventional methods with about 300 times faster processing speed than the NL-means filter.

Advances in ultrasonography: image formation and quality assessment

“…Also, the wide availability of GPUs enables implementations of image formation algorithms that require more intensive computations [98,99]. Such a research environment should further accelerate explorations in this field, and the deep neural network should also provide a powerful option for medical ultrasound beamforming [100][101][102][103][104], ultrasound image processing such as speckle reduction [105,106], and image segmentation [107,108], etc. Through such investigations, ultrasonography will increase its value in medical diagnostics.…”

Ultrasound Signal Processing: From Models to Deep Learning

“…Both Vedula et al (Vedula et al, 2017) and Ando et al (Ando et al, 2020) approached the issue of speckle reduction in similar ways, by using a CNN. However, they used different forms of input and target data, while Vedula et al (Vedula et al, 2017) used IQ data, Ando et al (Ando et al, 2020) used B-mode images.…”

Ultrasound Signal Processing: From Models to Deep Learning

“…The authors achieve this by training a deep CNN with a dataset containing B-mode US images across a range of scale/zoom factors. A similar approach was taken byChoi et al (2018), where they propose a deep CNN called SRGAN with the aim to map low resolution images to a high resolution domain.BothVedula et al (2017) andAndo et al (2020) approach the issue of speckle reduction is similar ways, by using a CNN. However, they use different forms of input and target data, whileVedula et al (2017) use IQ data,Ando et al (2020) uses B-mode images.…”

“…A similar approach was taken byChoi et al (2018), where they propose a deep CNN called SRGAN with the aim to map low resolution images to a high resolution domain.BothVedula et al (2017) andAndo et al (2020) approach the issue of speckle reduction is similar ways, by using a CNN. However, they use different forms of input and target data, whileVedula et al (2017) use IQ data,Ando et al (2020) uses B-mode images. Similarly,Dietrichson et al (2018) perform this task by using a CNN, albeit with a more complex training strategy, by employing a GAN based structure Karaoglu et al (2021).…”

“…While all the works cited on this topic so far deal with 2D US scans,Li et al (2020) attempt to extend it to 3D imaging using a 3D UNet model. It is an interesting point to note thatVedula et al (2017) andAndo et al (2020) use simulated data, while the other works on speckle reduction use in-vivo data gathered from volunteers. However, there is uniformity in how these works create their target images; through model based speckle reduction algorithms.Most deep learning methods for image quality improvement rely on supervised learning, requiring ground truth targets which are often difficult to obtain.…”