Medical Image Registration Aligns Scans

Registration warps moving scans onto fixed ones so the same voxel means the same tissue across time and machines.

Why Does This Exist?

Tumor growth compares January against June, but patients pose differently and breathe. Subtraction without alignment paints motion as growth. Registration optimizes transform $T$ minimizing dissimilarity plus regularization. Rigid covers head frames; deformable covers breathing livers. Tumor segmentation then measures on aligned grids.

Think of It Like This

Aligning tracing paper across semesters of maps

You overlay this semester's campus map onto last year's by pinning two halls, then stretching the paper where new wings bulge. Brick halls pin rigidly while lawns stretch. Measuring new construction needs the pinned stack.

Fixed scans are old maps and warps are stretches. The analogy stops at folding: paper can crease while tissue warps stay invertible.

How It Actually Works

Rigid $T$ holds $6$ numbers (rotation plus translation); affine adds $12$ ; deformable holds a displacement per voxel regularized by bending energy. Mutual information suits CT to MRI since intensities differ. Target error under $2$ mm suits tumors; lungs accept $3$ under breathing.

Worked example

A liver lesion sits at $(120, 80, 45)$ in January and $(122, 81, 47)$ in June from pose plus $3$ mm growth. Rigid alignment removes $(2, 1, 1)$ pose, leaving $(0, 0, 1)$ plus growth. Deformable with $5$ mm smoothing keeps vessels aligned within $1.5$ mm while the lesion grows $4$ mm. Without it, growth reads $7$ mm and mistimes care.

Watch Out For

Deformable warps erasing real growth

Over-regularized warps squeeze growing tumors back to baseline to please similarity. The symptom is stable disease that later bursts. Cap deformable magnitude near tumors and measure on rigid plus local review.

The Quick Version

Registration aligns scans so voxels compare across time.
Rigid suits heads; deformable suits breathing abdomens.
Mutual information bridges CT and MRI intensity gaps.
Do not let warps squeeze tumors away; cap near lesions.
Two-millimeter error suits tumors; lungs tolerate more.