Southeast University's VCC-DSA Pushes Stroke Imaging Quality up 73% Over Deep Learning
A vascular consistency constrained model cleans up motion artifacts in cerebral angiography, hitting 43.2 dB PSNR and 98.8% SSIM on animal ground truth.
ProTakeaways
- VCC-DSA published in Medical Image Analysis tackles motion artifacts in cerebral angiography.
- Two synchronous forward passes with shared weights force vessel output to stay consistent across different mask frames.
- Achieves 98.8% SSIM and 43.19 dB PSNR on motion-free anesthetized pig ground truth, 82 ms per frame.
- Beats Gao 2019, Yonezawa 2022 pix2pix, and U-Net baselines by roughly 18 dB PSNR under fair evaluation.
- Mixup-based Data Self-evolution builds a Vascular Bank that iteratively cleans contaminated training labels.
- Fails when eyeball rotation mechanically drags peripheral vessels, breaking the consistency assumption.
Digital Subtraction Angiography is the gold standard for spotting stroke-causing clots and aneurysms, but it has a stubborn flaw. Patients breathe, twitch, and their hearts beat, so bones and teeth end up in slightly different positions between the two frames doctors subtract. The residue looks like fake blood vessels, and it lands right on top of the real ones. A team from Southeast University, United Imaging Healthcare, and Case Western has published VCC-DSA in Medical Image Analysis, a learning-based imaging model that pushes PSNR by 73.4% and SSIM by 8.56% over prior deep learning baselines on motion-free animal ground truth.
Why subtracting two X-rays is harder than it sounds
Clinical DSA often suffers from misregistration artifact caused by voluntary, respiratory, and cardiac motion during acquisition. The classical trick is simple: take a mask image before contrast agent is injected, take a live image after, subtract the two, and the vasculature lights up while everything else vanishes. In practice, teeth, skull edges, oxygen tubes, and catheters barely move but not exactly, and the leftover ghosts smear over the anatomy of interest.

The field has tried two workarounds, both leaky:
- Registration-based methods warp the mask to match the live image before subtracting. These approaches compensate for motion effectively but require time-consuming iterative computation for each frame, which limits clinical applicability. They also fight a losing battle, since patient motion is 3D while the X-ray projection is 2D, so no planar warp can fully recover the misalignment.
- Style-transfer synthesis throws away the mask and trains a network (typically pix2pix or a GAN) to hallucinate a DSA image straight from the live frame. These models do not explicitly compensate for motion-induced misalignment, often hallucinate or modify contrast and vessels, and lack interpretability. Bone that looks tubular gets painted in as a fake artery; real vessels overlapping bone get erased.