The algorithm in this paper was developed solely using the grayscale images without any knowledge of the ultrasound signal processing path inside the system. We have thus demonstrated that the automated angle estimation technique can be applied on ultrasound images after pre-processing steps have been applied. Such an approach makes the algorithm agnostic to the ultrasound pre-processing and optimization algorithms that are typically applied before image formation in an ultrasound system. As a result, this opens the possibility that the algorithm can be applied as an add-on module on different ultrasound system models and those manufactured by various vendors. A proposed next step in this study would be to include images from multiple ultrasound systems and vendors in the data set. Note that in this study, all of the images were acquired by a single ultrasound system. This extension would help test the robustness of the technique and make the approach more broadly applicable.
The data set used in this paper consisted of multiple images for a given patient, each with a different Doppler angle, to augment the data set. An extension of this work would be to test the algorithm on a larger data set that comprises images from multiple patients. Given the anatomical variations between patients that results in a range of vessel orientation, this would further evaluate the robustness of the approach on patients with a large range of Doppler angles.
Conclusion
To the best of the knowledge of the authors, this is the first publication demonstrating the use of deep learning for automated angle estimation in Doppler ultrasound. Previous approaches relied on computer vision approaches where image features were manually extracted using established techniques such as segmentation. The novel approach proposed in this paper has the potential to significantly reduce the examination time for performing a Doppler exam, and thus making the clinical ultrasound workflow more efficient.