SkullGAN Filtered Synthetic Skulls

To address the challenges of limited and heterogeneous skull CT data for training deep learning models in transcranial ultrasound (TUS) applications, we developed SkullGAN, a generative adversarial network designed to generate high-quality synthetic skull CT slices. SkullGAN leverages real CT data from 38 healthy subjects to produce synthetic images that closely mimic the radiological features of real skulls, as evaluated through skull density ratio, mean thickness, mean intensity, t-SNE analysis, a Fréchet inception distance (FID) score of 49, and a visual Turing test (VTT) with expert radiologists achieving 60% mean accuracy in distinguishing real from synthetic images. These synthetic skull slices were utilized to train TUSNet, a novel deep learning framework for end-to-end transcranial ultrasound simulation and phase aberration correction. TUSNet achieves unparalleled efficiency and accuracy, computing ultrasound pressure fields and phase corrections in 21 milliseconds (over 1200× faster than k-Wave) with a mean focal positioning error of 0.18 mm and 98.3% peak pressure accuracy. This dataset includes the 5,000 synthetic skull CTs that we sampled from SkullGAN for our training, along with indices of similar skulls that should be excluded to yield the TUSNet training set. More details about this filtering are available on the TUSNet Github repository. By publishing this data (and the SkullGAN model), we aim to advance TUS research by providing a scalable and accessible resource for training and validating deep learning models.