MA
Portfolio

Area BProduction ML Platform

Series Registration

Novel optimization · from-scratch implementation

Aligning prior and current scans from three landmark pairs, ~4× faster than the library baseline.

target registration error
~0.41 mm
target registration error
faster than volume registration
~4×
faster than volume registration

01/Problem

Radiologists compare a new scan against a prior acquired at a different time, orientation and sometimes scanner. Intensity-based registration is accurate but heavy and needs per-modality tuning.

02/Approach

Volume registration on SimpleITK, plus a from-scratch landmark optimizer that aligns two studies from just three radiologist-provided landmark pairs — a 7-DOF similarity transform solved by gradient descent with hand-derived analytical gradients for all seven parameters, including the quaternion coupling. It needs no voxel data at inference, making it modality- and anatomy-independent.

03/Outcome

  • ~0.41 mm average target registration error across 58 multi-modal series pairs in ~0.9 s.
  • Roughly 4× faster than SimpleITK volume registration (~4 s).
  • Volume registration benchmarked over 20 repeated runs with 95% CIs across MR head, CT head and CT lung.

04/My role

  • Approved a from-scratch optimizer over tuning the library path, judging modality-independence and sub-second latency worth the build.
  • Required repeated trials with confidence intervals rather than single-run timings.

Stack

NumPySimpleITKElastixQuaternions