Transcranial focused ultrasound stimulation (tFUS) is a promising noninvasive neuromodulation technique that offers high spatial resolution compared to other non-invasive alternatives. However, its effectiveness is hindered by skull-induced phase aberrations, which distort the ultrasound beam and reduce focal accuracy. Existing solutions are often computationally intensive or fail to account for the effects of ultrasound waves outside the expected focal spot.

In this paper, we approach phase correction as a model inversion problem and propose the Dual-Branch Skull-Induced Phase Aberration Correction Network (DB-SIPAC), a novel domain-enriched machine learning framework designed to efficiently predict time-delay profiles. The DB-SIPAC model consists of two specialized branches: the Pathway Branch, which focuses on the direct path from the ultrasound transducer to the target, and the Skull Branch, which incorporates full skull structure information to account for reflections and refractions. This dual-branch design enables rapid and accurate time-delay predictions, reducing computational time from hours to less than a second. To facilitate learning, we generated a training dataset consisting of skull images and ground-truth time-delay profiles across a range of skull shapes and thicknesses. Since existing methods cannot guarantee time-delay profiles that ensure the expected focal spot has the maximum pressure within the brain, we introduce the Iterative Time Delay Search (ITDS), a novel numerical algorithm that iteratively refines time-delay profiles generated by state-of-the-art methods to generate the ground truth for DB-SIPAC. Simulation results demonstrate that DB-SIPAC outperforms state-of-the-art alternatives, achieving perfect focal point alignment with the target and maximizing focal pressure, all while providing real-time inference.