Unsupervised deep learning for ultrasound beamforming and beyond

点击翻译按钮获取中文摘要

Even though ultrasound is the second most popular medical imaging modality, its potential is far greater. Specifically, channel radio-frequency (RF) data is not suitable for visualization and, as such, is beamformed to the familiar gray-scale B-mode (brightness mode) ultrasound images. This conversion, however, is lossy and destroys most of the information in the RF data. Recent advances in AI and ultrasound provide a perfect combination for taking advantage of this trove of discarded information to further widen ultrasound use. A critical challenge in exploiting advances in deep learning (DL) in ultrasound beamforming lies in a lack of training data in real experiments. More specifically, the ground truth or optimal image quality in real experiments is unknown; therefore, simulations are performed to train deep models. However, simulations always make simplifying assumptions and, as such, do not perfectly model real-world ultrasound data. This collaboration has two main aims to address this issue:Aim 1. Unsupervised fine-tuning on real ultrasound data by introducing auxiliary tasks.Aim 2. Beyond conventional beamforming by unified deconvolution and beamforming with DL.This project will be done in collaboration with the CREATIS lab, which is one of the main French labs in medical imaging with a workforce of >200 members. The added benefits of the collaboration are three-fold. First, the Canadian and French teams provide complementary expertise. Second, France and Canada are home to several industry leaders in ultrasound. Therefore, this collaboration facilitates technology transfer from academia to industry. And third, this project relies heavily on a database provided by the CREATIS lab; hence, the collaborator's expertise is critical in processing the data and analyzing the results.