Enhanced Musculoskeletal Ultrasound Imaging through PatchGAN Adaptation of Raw Channel Data

Deep Learning (DL) based accelerators for beamforming have been gaining popularity as an ideal candidate to enhance image quality for point-of-care ultrasound (POCUS) devices in low-resource settings. POCUS devices are widely used for musculoskeletal (MSK) imaging, particularly for assessing bone structures. However, they often exhibit poor contrast, high speckle noise, and limited resolution, resulting in inadequate visualization of bone structures in MSK applications. In this work, we present a deep learning-based beamforming approach that reconstructs ultrasound bone images directly from radio frequency (RF) data, enabling accurate MSK assessment in POCUS systems. In this work, we present a deep learning-based beamforming approach that reconstructs ultrasound bone images directly from radio frequency (RF) data, learning both time-of-flight correction and apodization to enable accurate MSK assessment in POCUS systems. The model was trained and evaluated on in-vivo MSK datasets. The model achieved a mean CNR of 6.37 dB, SNR of 1.57 dB, and SSIM of 0.87, values comparable to those of the conventional DAS method (CNR: 6.32 dB, SNR: 1.52 dB). These findings suggest that the model effectively learns the beamforming process, can replicate the image quality of conventional DAS with lower computational requirements, and hence can be extended to learn adaptive beamforming algorithms.