Improving Image Quality of Single Plane Wave Ultrasound via Deep Learning Based Channel Compounding

Deep Learning for Ultrasound Image Formation: CUBDL Evaluation Framework and Open Datasets

“…We also observed that learning a single, intermediate step of the image formation pipeline (as pursued by Goudarzi et al [29] and Rothlübbers et al [28]) represents a more clearly defined transformation for the presented task, as opposed to attempts to learn the entire beamforming process (i.e., the approach taken by [19], [27], which suffered from overfitting to the training data, as shown in the presentation of all challenge results [14]). To briefly summarize for the context of this manuscript, overfitting manifested as an inability of the submitted networks to detect point targets in the unseen test data, recreate lesions from the unseen test data, or replicate the gradation from light to dark shown in Fig.…”

“…However, there were statistically significant differences (p<0.05) when comparing the mean 2 -log and ρ of the networks submitted Based on the first additive term of Eq. 9, the images created with 75 plane wave transmissions achieved the highest rank in a majority of cases (i.e., 1), followed by images created with DAS beamforming after a single plane wave transmission, then the network submitted by Goudarzi et al [29], then the network submitted by Rothlübbers et al [28]. Specifically, the average ranks for image quality (i.e., the first additive term of Eq.…”

“…Details about their networks and performance results are available in [19], [27] and in the presentation document published on the CUBDL website [14], which was initially prepared for the live CUBDL challenge session on September 11, 2020 at the 2020 IEEE International Ultrasonics Symposium. The remaining two network submissions [28], [29] are compared and evaluated in greater detail in this publication, which upholds the commitment advertised on the CUBDL website that only top submissions would be included in the journal publication associated with the challenge [14]. Also as promised on the CUBDL website, the trained model weights for these top submissions are available with the public release of the CUBDL evaluation code [15] and datasets [16].…”

“…1) Rothlübbers et al [28]: A fully convolutional network was submitted by Rothlübbers et al [28], and the following summary includes details that are not available in [28]. The network had 4 layers.…”

“…OF LOCAL IMAGE QUALITY COMPARISONS AND NETWORK COMPLEXITY, WITH[28] AND[29] EVALUATED ON SINGLE PLANE WAVE DATA. Single Plane Wave -10.50 ± 5.40 1.05 ± 0.44 0.67 ± 0.23 1.85 ± 0.04 0.42 ± 0.07 0.36 ± 0.15 0 Rothlübbers et al [28] -9.82 ± 4.93 0.67 ± 0.26 0.53 ± 0.20 1.19 ± 0.16 0.32 ± 0.10 0.33 ± 0.13 3, 059 Goudarzi et al…”