Transcranial ultrasound localization microscopy (ULM) is a promising brain imaging modality, but phase aberration (PA) from the skull’s complex structure and acoustic impedance mismatch limits its performance. Traditional PA correction (PAC) methods like fast marching method (FMM) depend on accurate speed of sound (SOS) maps, which may not be available. Full wave inversion (FWI) techniques are SOS-independent but computationally intensive, making them impractical for ULM’s large data volumes. Deep learning (DL)-based PAC is efficient and model-free but requires realistic training datasets, which are difficult to obtain for in vivo applications. To address this, we propose a novel DL-PAC training strategy using paired microbubble data. Microbubble point spread functions (PSFs) were experimentally acquired by scanning diluted microbubbles in saline with and without excised mouse skulls, creating “banks” of aberrated and non-aberrated data. A lightweight residual convolutional network with adversarial training was used to map aberrated signals to their corrected counterparts. Experimental results show our DL-PAC significantly improved microbubble detection, generating 3.63 times more tracks under identical reconstruction conditions. The resulting ULM images displayed enhanced quality, with clearer microvessel delineation and improved recovery under skull-induced shadowing. The method was effective for mice up to 16 weeks old using 15-MHz imaging frequency, beyond which attenuation dominated.
Pengfei Song (Wed,) studied this question.