Deep learning approaches for quantitative analysis of ultrasound backscattered signals

Quantitative ultrasound (QUS) aims to improve diagnostic ultrasound imaging by extracting objective tissue parameters from backscattered signals. Deep learning can facilitate this process because of its ability to extract high-level information from the raw data. We have demonstrated that deep learning approaches applied to backscattered signals can accurately quantify liver fat noninvasively. We will discuss three deep learning approaches herein: (1) a one-dimensional convolutional neural network (1-D-CNN) for uncalibrated radiofrequency (RF) data; (2) a two-dimensional CNN (2-D-CNN) for uncalibrated spectrograms; and (3) a 2-D-CNN for calibrated spectrograms. Approaches (1) and (2) do not require system calibration with a physical phantom, whereas Approach (3) obtains calibrated spectrograms with a physical phantom. Each approach is evaluated using three datasets that contain ultrasound RF data acquired from the right liver lobe and contemporaneous MRI-PDFF (reference standard) in participants with and without nonalcoholic fatty liver disease: 204 scanned by a Siemens S2000 ultrasound scanner (4C1), 70 scanned by a Siemens S3000 scanner (4C1 and/or 6C1HD), and 104 scanned by both the S3000 and a GE Logiq e9 scanner (C1-6). We will discuss the advantages and disadvantages of each approach evaluated using data acquired across multiple ultrasound scanner platforms. [Supported by R01DK106419.]Quantitative ultrasound (QUS) aims to improve diagnostic ultrasound imaging by extracting objective tissue parameters from backscattered signals. Deep learning can facilitate this process because of its ability to extract high-level information from the raw data. We have demonstrated that deep learning approaches applied to backscattered signals can accurately quantify liver fat noninvasively. We will discuss three deep learning approaches herein: (1) a one-dimensional convolutional neural network (1-D-CNN) for uncalibrated radiofrequency (RF) data; (2) a two-dimensional CNN (2-D-CNN) for uncalibrated spectrograms; and (3) a 2-D-CNN for calibrated spectrograms. Approaches (1) and (2) do not require system calibration with a physical phantom, whereas Approach (3) obtains calibrated spectrograms with a physical phantom. Each approach is evaluated using three datasets that contain ultrasound RF data acquired from the right liver lobe and contemporaneous MRI-PDFF (reference standard) in participants with and ...