Research

Research Assistant, LDR Group · Johns Hopkins University

Advised by Prof. Peirong Liu (ECE / DSAI) and Risheng Xu, M.D., Ph.D. (Neurosurgery)

• Reframed MRI–TOF-MRA fusion for trigeminal neuralgia as an ROI-centered neurovascular evaluation problem rather than generic whole-brain registration.
• Built the end-to-end preprocessing, registration, and evaluation pipeline over a 149-patient microvascular-decompression cohort with 298 clinician-annotated trigeminal ROIs.
• Benchmarked six registration pipelines (ANTs Affine/SyN, ConvexAdam, FireANTs, EasyReg, SynthMorph) and showed that image-similarity, vessel-separability, and downstream localization metrics do not co-rank methods.
• Identified a volume confound in the standard one-sided vessel distance (Spearman ρ = −0.53, versus +0.08 in the reverse direction): common metrics systematically favor methods producing larger vessel trees under partial clinical annotations.
• Designed and ran a blinded two-neurosurgeon reader study over 100 stratified ROIs, exposing globally implausible registrations hidden by favorable local metrics — the best-scoring method failed at the whole-brain level in 71 of 100 cases.

X. Zhang, X. Wang, M. Xie, H. Liang, H. E. Lien, O. Das, J. Feghali, R. Xu, P. Liu. “Evaluation Principles for MRI–MRA Registration in Trigeminal Neuralgia: An ROI-Centered Neurovascular Benchmark.” IEEE Transactions on Medical Imaging, under review, 2026. The benchmark and evaluation code will be released publicly.

Independent Research Project, LDR Group · Johns Hopkins University

Advised by Prof. Peirong Liu (ECE / DSAI)

• Investigated whether cerebral vessels can be segmented from routinely acquired T2 MRI alone, using paired MRA only as a training-time supervision source — removing the MRA dependency that constrains the work above.
• Repurposed the VesselFM foundation model as an MRA-side teacher — it collapses when applied directly to T2 (Dice 0.041) — warping its pseudo-labels into T2 space to supervise a 3D nnU-Net student that requires no MRA at inference.
• Reached Dice 0.489 ± 0.043 on held-out subjects from a 568-subject IXI cohort — +0.448 over zero-shot VesselFM and within 0.072 of an oracle trained on manual MRA labels, using no manual annotation.
• Separated label quality from label quantity: at identical training-set size, teacher-derived pseudo-labels beat placeholder labels by +0.118 Dice, showing that pseudo-label scale helps only when the teacher observes the correct modality.

• Co-led evaluation for Task 4 (Trigeminal Neuralgia Segmentation): multiclass segmentation of the trigeminal nerve and surrounding vasculature from T2-weighted MRI, scored by DSC and NSD.