Applied ML / Research workbench / Engineering notes
MotionTrace
A research workbench for tracking an initially annotated target through prerecorded cine-MRI and examining reliability.
Python
A small descriptive analysis found the learned candidate underperformed causal ZNCC on Dice and centroid error in five of six evaluation patients. No clinical validation.
The work
A research workbench for tracking an initially annotated target through prerecorded cine-MRI and examining reliability.
Evidence
A small descriptive analysis found the learned candidate underperformed causal ZNCC on Dice and centroid error in five of six evaluation patients. No clinical validation.
Scope
Research and education prototype; not clinical software. Evaluation data are exposed, and publication rights for research assets remain pending.
The decision, at a glance
Did the learned method improve tracking?
- 01
Initialize a target
Begin from an annotated target in a prerecorded sequence.
- 02
Compare methods
Inspect fixed-mask reference, causal ZNCC and a learned candidate.
- 03
Preserve the finding
ZNCC outperformed the learned candidate on Dice and centroid error in five of six evaluation patients.
A small descriptive comparison on exposed evaluation data; the negative result is retained.
Read the verification scope ↗Original text-only diagram. No patient imagery, weights or dataset-derived media. Not clinical software.
The problem
Investigate tracking an initially annotated target through prerecorded cine-MRI, and make method behavior and limitations easier to inspect.
Implementation
I developed a local research workbench and compared a fixed-mask reference, a causal ZNCC tracker and a bounded learned-model candidate.
One decision: retain the negative result
The learned candidate underperformed ZNCC on Dice and centroid error in five of six evaluation patients in the corrected analysis. This descriptive finding is preserved rather than presented as a model improvement.
Demonstration
The workbench and portfolio package are local. Research images, weights and dataset-derived assets are withheld pending publication rights and exact-file approval.
Evidence and limitations
The study is small, evaluation data have been exposed, and the work is not clinically validated. It is not diagnosis, autonomous tumor discovery, treatment planning or equipment control.