Toward ultrasound brain imaging via machine-learning-extracted skull profile and speed of sound

Project Summary Transcranial ultrasound could facilitate a broad variety of applications in brain imaging, e.g., functional imaging, intracerebral hemorrhage detection, brain perfusion evaluation, and stroke diagnosis. Ultrasound has the intrinsic advantages of being real-time, portable, widely available, noninvasive, and free from ionizing radiation. Thus,

transcranial ultrasound imaging could potentially play a unique role in a time-sensitive and dynamic environment where X-ray computed tomography (CT) and magnetic resonance imaging (MRI) are unavailable. For instance, transcranial ultrasound has significant potentials in the initial assessment of traumatic brain injury during the

transportation of patients to the hospital, and in bedside monitoring of brain physiology for stroke patients in an intensive care unit. Despite its promising potentials, the use of transcranial ultrasound imaging has been limited, largely because adult human skulls cause severe phase aberration, leading to highly degraded ultrasound

images. Phase aberration from the skull can be accurately corrected if the speed of sound (SOS) and profile (i.e., thickness distribution) of the skull are known a priori. The skull profile and SOS can be estimated by CT, currently the gold standard approach for treatment planning. The CT-based approach is far less appealing,

however, for ultrasound imaging purposes because of the additional CT scans that involve ionizing radiation and image co-registration. We propose a real-time pulse-echo ultrasound approach to estimate the skull profile and SOS using deep learning (DL) methods with ultrasound radiofrequency (RF) signals backscattered from the skull.

The proposed approach rests on the scientific premise that these RF signals contain extremely rich information of the interaction between ultrasound and skulls, and the information of skull profile and SOS is encoded in the backscattered signals in a convoluted way that cannot be fully described by simple physical models. We

hypothesize that DL, a subclass of machine learning (ML), is capable of automatically and rapidly extracting skull profile and SOS from RF signals with sufficient training. The objective of this Trailblazer R21 application is to develop and validate DL methods for extracting the human skull profile and SOS, with the following aims. Aim 1.

In silico study: Develop and evaluate DL-based skull profile and SOS extraction algorithms using synthetic data. Aim 2. Experimental study: Evaluate DL algorithms’ performance in skull profile and SOS extraction using experimental data. Aim 3. Pilot imaging study: Evaluate DL algorithms’ performance in transcranial imaging.

Successful completion of this study will facilitate the transcranial application of both conventional (e.g., B-mode imaging, blood flow imaging, and contrast-enhanced ultrasound) and emerging ultrasound imaging methods (e.g, super-resolution imaging and photoacoustic tomography). Although the current application focuses on brain

imaging, our method can be extended to phase aberration correction for ultrasound-based brain treatment, neuromodulation, and ultrasound imaging of other organs where phase aberration exists.