Background & Limitation: Ultrasound signals become blurred and distorted as they pass through the skull, making it difficult to clearly visualize cerebral vasculature. A major bottleneck is the absence of ground-truth open-skull and transcranial observation pairs.
Our Goal: To overcome this, we generate structure-preserving pseudo pairs and train a model to restore transcranial images to open-skull quality.
MIST: We propose a multi-stage network. First, CycleGAN with Localization Loss generates synthetic paired data. Second, a Uformer model restores these images, performing denoising, resolution enhancement, and PSF restoration.
We utilize CycleGAN with a novel Localization Loss to convert open-skull images into transcranial-style images while preserving microbubble dynamics.
A Uformer-based supervised model learns to reverse the degradation (Open-skull ↔ Synthetic Transcranial) to remove skull-induced artifacts.
Quantitative: MIST achieves significantly higher PSNR compared to baselines (Richardson-Lucy, CycleGAN-only), improving by 7-8 dB.
Qualitative: Restores faint vessels and sharpens the image, effectively overcoming skull-induced blur.
MIST restores severely degraded transcranial ultrasound into clear cerebrovascular images without invasive craniotomy. It delivers vascular enhancement and significant PSNR gains. Future work will expand evaluations to verify physical fidelity.
Jihyeok Jung is an undergraduate researcher in Electronic Engineering at Sogang University. His primary research focuses on Multimodal and Vision-Language Models, leading to a first-author paper at CVPR 2025 on egocentric instruction tuning. His research interests also actively extend to the medical domain, particularly deep learning for ultrasound imaging.
Hyungmin Lee is a researcher at the Sogang MIST Laboratory, specializing in developing algorithms for transcranial ultrasound restoration. His work emphasizes artifact suppression and image quality enhancement in challenging scenarios.
