Complex-Valued Symmetric Series Convolutional Network for Ultrasound Super-Resolution Image Reconstruction

This work presents a deep learning architecture designed to enhance ultrasound image reconstruction by utilizing complex-valued ultrasound data. Specifically, it proposes a complex-valued extension of the symmetric series convolutional neural network originally built for super-resolution image reconstruction from image data. The model leverages a complex-valued feature extraction to preserve phase information and employs a tailored upsampling strategy based on sub-pixel convolution. Both the complex and real-valued versions of the model were trained on simulated ultrasound dataset and assessed using synthetic and in-vitro phantom data. Results from the evaluation reveal that the proposed approach successfully reconstructs high-resolution ultrasound images from a reduced number of acquisitions, surpassing traditional methods.