Measuring how a tumour changes, consistently, between scans.

The problem

Judging whether a tumour has grown means comparing two scans taken weeks or months apart, slice by slice. It's slow, and it's subjective: two experienced readers looking at the same pair of studies can reach meaningfully different conclusions about how much has changed, because the human eye estimates volume badly.

What clinicians and researchers wanted was not an opinion from a machine. It was a number, produced the same way every time.

The value wasn’t speed. It was consistency — the same input producing the same measurement, every time, for every reader.

What we built

A deep-learning pipeline that takes clinical imaging from raw format through to a comparative report.

• Format normalisation. DICOM studies converted to a research-standard volumetric format, with orientation and spacing handled correctly rather than assumed.
• Segmentation. A transformer-based segmentation model identifies three distinct tumour sub-regions — necrotic core, oedema and invasion, and enhancing tumour — rather than treating the mass as one undifferentiated blob.
• Volumetric quantification. Each region measured, and the change between timepoints calculated and reported numerically.
• Reporting both ways. Visual overlays for the reader, quantitative tables for the record, and results exportable back into DICOM so they live where clinicians already work.

How the agents were used

The segmentation, quantification and reporting stack is the AIOBC Vision Agent, in its clinical configuration. Nothing about this pipeline depended on a third-party API — it was built entirely on in-house and open-source components, which for medical imaging is a requirement rather than a preference. Data never leaves the environment it belongs in.

The outcome

Comparison between studies became a quantified, reproducible process rather than a visual estimate. Readers get the measurement and the overlay; the record gets a number that means the same thing next month as it does today.

A necessary word about clinical AI

This is a measurement and research tool. It does not diagnose, it does not stage disease, and it does not make clinical decisions — a qualified clinician does, and the system exists to give them a better instrument, not to replace their judgement. Any deployment into patient care carries regulatory obligations that sit well beyond the software itself, and we say so early rather than late.