Back to Bigger Picture

BIGGER PICTURE Earlier diagnosis

AI found 26 pancreatic lesions that initial care had missed

A CT taken for one reason may contain clues to another disease. PANDA explored whether routinely acquired scans without contrast could reveal pancreatic abnormalities that otherwise received no focused attention.

AI-generated illustration: A scan-analysis concept highlights a small pancreatic finding for review.
A scan-analysis concept highlights a small pancreatic finding for review.AI-generated illustration of the idea, not a photograph of the actual people, equipment or event.
Real-world evaluation
20,530 consecutive patients
Initially missed findings
26 pancreatic or nearby lesions
Of those findings
One pancreatic ductal adenocarcinoma

How the idea works

  1. Reuse an existing CT

    Non-contrast images include the pancreas region.

  2. Locate and classify tissue

    The model identifies lesions using learned anatomical and diagnostic patterns.

  3. Investigate the flag

    Clinicians decide what additional assessment a finding needs.

Use information already captured

Non-contrast abdominal scans, and some chest scans, include part or all of the pancreas. The challenge is that pancreatic lesions can be difficult to distinguish without contrast. PANDA was designed to inspect those existing images, avoiding an additional scan solely for its initial analysis.

Sources: [1]

Use clearer images to teach a harder task

During training, expert markings from paired contrast-enhanced scans were transferred onto non-contrast images. Pathology-confirmed lesion types supplied another reference. The model learned to locate the pancreas and suspicious tissue, then distinguish lesion categories rather than merely label an entire scan abnormal.

Sources: [1]

What an extra look uncovered

The 2023 paper included large retrospective evaluations in routine clinical settings. PANDA identified twenty-six lesions missed in the initial care assessment. They included one pancreatic ductal adenocarcinoma, other tumours, cysts and inflammatory findings. Some had already been discovered during later routine follow-up before the retrospective analysis.

Sources: [1]

Make the next investigation more targeted

A flagged image still needs clinical assessment and sometimes further imaging or tissue diagnosis. The study did not prove that population screening with PANDA extends life. Its compelling idea is opportunistic detection: extract another useful question from data already collected, then test whether acting on it benefits patients.

Sources: [1]

The original sources

  1. Nature Medicine / author manuscriptLarge-scale pancreatic cancer detection via non-contrast CT and deep learning

What to explore next

The X-ray clinic travelled to the minersThe feeding robot met its hardest test at home
Explore all 99 applications