More input isn't better input
Giving a model more information during training didn't sharpen its edges on its own. How information shapes learning matters more than how much of it there is.
Research
Teaching a model where the edge really is.
We're building segmentation that follows a tumour's true boundary and is honest about where it's unsure. Here's what we've learned so far, including what didn't work.
The problem
Finding a tumour isn't the same as outlining it.
Pancreatic cancer is infiltrative. It wraps vessels, blends into surrounding tissue and often extends past the gland itself. A useful outline has to show where it really ends.
Most segmentation pipelines are tuned to find tumours. They approximate the shape, then trim anything outside the organ. That's fine for detection and wrong at exactly the edge a radiologist cares about.
We change how the model learns, not what it needs at the point of use. CT in, soft contour out: no report, no extra inputs.
What we've learned
Four lessons, measured.
Each comes from a controlled internal study on development scans. They describe the problem; they are not claims about clinical performance.
Lesson 1
Rigid anatomy rules cost real tumour
Share of expert-labelled tumour that falls outside the region a standard pipeline lets the model search. 143 training cases.
One in seven tumour voxels can sit outside the search region. In an earlier check, 7 of 8 tumours extended past the organ outline, and a strict cut would have deleted 27.9 mL of correctly found tumour to remove 1.7 mL of false alarms.
That trade is backwards for delineation, and it's the core problem Soft Contour is built to fix.
Lesson 2
Loosening the rules isn't enough
The obvious fix, simply letting the model look wider, barely moved overlap. It was far below the gain we set in advance, and it added roughly 80 voxels of wrong tissue for every voxel of tumour recovered. The edge needs a smarter signal, not a bigger search.
Change in overlap from a wider search
Local fixes can hurt the whole
Several changes corrected errors where we looked and made the full outline worse. Every change we ship is judged on the whole scan, not a flattering slice.
How we work
Every claim earns its place.
Medical AI loses trust fastest when it overclaims. So we hold ourselves to rules set before we look at results.
- Beat a fair comparison. A change counts only if it beats both the current model and a matched control.
- No hidden costs. Better edges can't come with more missed tumour or more false alarms.
- Reproducible. Results hold across repeated training runs, with every case kept.
- Nulls count. What didn't work shapes what we build next, so we keep it.
- CT only at the point of use. Nothing extra for the radiologist to supply.
See it working
Soft contours on real CT.
The live demo runs the product on real scans. Scroll through the slices and compare the soft contour with a hard mask.