Shevaan Jayasinghe
← More work

Pancreatic tumour segmentation · Research

A description isn’t
an outline.

A radiology report can say where a tumour is and how large it looks without drawing its edges. Our research asks how that incomplete information can help a model learn what to look for in a scan.

Problem
A report can say where a tumour is without drawing its edges, and a search limited to the pancreas could miss tumour beyond it.
Role
With a teammate, I examined what that search missed, and I worked on the Apple GPU path for a calculation that was hitting its time limit.
Result
The wider search recovered missed tumour but admitted more incorrect tissue, so we didn’t advance it.
REPORTWhere it isHow large it looksIts edgesA description,not an outline.SCANMISSEDRECOVEREDLET INSEARCHNARROWWIDERREPORTWhere it isHow large it looksIts edgesA description, not an outline.SCANMISSEDRECOVEREDLET INSEARCHNARROWWIDER

Concept illustration

01

The question inside the research

What if the search is missing part of the answer?

Working with a teammate on anatomy and error analysis, I examined whether a search restricted by the pancreas could leave out tumour that extended beyond it. It did in some of the training scans we checked.

02

The result that changed the direction

More of the right tissue. Too much of the wrong tissue.

A large calculation was hitting its time limit, so I worked on the Apple GPU path, narrowing the candidate search and refining it in small pieces while keeping the target-construction rules the same. The full comparison then ran: the softened search recovered missed tumour but also admitted substantially more incorrect tissue, with no clear improvement in the quality of the training targets, so we did not advance that candidate.

For a closer look at the engineering

The search-region check found tumour outside the anatomy-derived restriction in 17 of 143 training scans. The follow-up compared hard, softened and unrestricted searches using the same saved model probabilities and size inputs.

The softened search recovered some missed tumour but admitted substantially more incorrect tissue. It showed no clear improvement in target quality, so that candidate was not advanced. These were comparisons of constructed training targets, not improved performance from a newly trained model on unseen scans.

For the GPU repair, FFT screening narrowed candidates and the original convolution refined them in small tiles. This bounded the large temporary allocation while retaining the downstream target-construction rules. Outputs matched the available completed reference cases; the repaired path allowed the full comparison to finish.

Live products

Open what is already shipped.

Open the live demo

Chappie is on the App Store. The AI Bubble Index is live on this site. This page is a record of earlier work.