Tumor Segmentation with Overlap Loss

Tumor segmentation separates active core from edema voxel by voxel because treatment targets the core, not the swelling.

Why Does This Exist?

Radiation boosts the enhancing core while sparing edema, so lumping them overdoses brain. MRI shows core on post-contrast T1 and edema on FLAIR. Input is four MRI sequences; output is three zones: whole, core, enhancing. Cores under $2$ mL punish pixel losses, which is why overlap losses lead.

Think of It Like This

Marking a peach bruise pit versus spread

A dropped peach has a brown pit plus a pale spread. Cooks cut the pit widely and the spread narrowly. Marking both as bruise wastes fruit.

Enhancing core is the pit and edema is the spread. The analogy stops at growth: tumors infiltrate while bruises heal.

How It Actually Works

A 3D net predicts whole, core, and enhancing masks jointly with Dice per zone plus cross-entropy. Enhancing needs post-gadolinium T1; missing it drops that zone to guesswork. Margins expand the core $5$ mm for planning but metrics score the raw mask.

Worked example

A glioblastoma spans whole $42$ mL, core $11$ mL, enhancing $6$ mL. The net predicts enhancing $6.5$ mL with $5.4$ overlap, Dice $2 \times 5.4 / 12.5 \approx 0.86$ . Edema predicts $31$ mL with $27$ overlap, Dice $0.88$ . A $2$ mm rim shift drops enhancing Dice $0.12$ but whole only $0.03$ , so rims need expert review.

Watch Out For

Blood products mimicking enhancement

Subacute blood glows on pre-contrast T1, so subtracting wrong phases fakes enhancing tumor. The symptom is enhancing that never washes. Compare pre against post before labeling; true enhancement appears only post.

The Quick Version

Tumors split into whole, core, and enhancing zones for dosing.
Post-contrast T1 shows enhancing rims; FLAIR shows edema spread.
Dice per zone beats pixel accuracy on tiny cores.
Do not plan on edema alone; boost the enhancing core.
Blood can fake enhancement without pre-post comparison.