A pixel's displacement is set almost entirely by its immediate neighbourhood — ultrasound displacement is a fundamentally local problem. NCA-Flow takes that literally: one small update rule, applied identically to every cell and iterated twenty times, until a dense displacement field emerges from nothing.
The work spans the whole pipeline: acoustic simulation for supervision, a dense annotation instrument for real clips, the architecture and its ablations, anchored sim-to-real transfer, and a four-metric evaluation — all owned end to end.
trainable parameters
pooled target-registration error
smaller than StrainNet-f — and more accurate
field folding, against StrainNet-f’s 31.23%
Final thesis numbers · pooled over 3 subject-disjoint clips
Estimate dense tissue displacement in B-mode ultrasound at every pixel, not at a handful of tracked landmarks.
Master thesis researcher owning acoustic simulation, annotation tooling, architecture, training, and evaluation end to end.
PyMUST compression simulation, a DUSTrack annotation wrapper, NCA-Flow training, anchored sim-to-real transfer, SLURM runs.
Dense vector regression by iterated local updates, a fixed differential perception basis, variance-regularised L1, leakage-aware subject-disjoint evaluation.
0.674 mm pooled TRE at 16,714 parameters — more accurate than every learned baseline, at between 40× and 2,314× smaller.
Ex-vivo and phantom only. No in-vivo data, no clinical claim. Submitted to MICCAI 2026 and not accepted.
Local iterative updates gradually refine a dense tissue-displacement field. Teal indicates compression; amber indicates expansion.