An image score from routine H&E added prognostic information in resected pancreatic cancer

Design

Multicohort development and validation, including prospectively accrued cohort with offline inference

Population

644 training cases and six test cohorts (1,443 patients, 155 of them prospective) with resected pancreatic ductal adenocarcinoma

Primary outcome

Overall survival

Effect

Prospective cohort: mortality 47.5% vs 22.7%; adjusted HR 4.45 (95% CI 2.28 to 8.70)

Patients with resected pancreatic ductal adenocarcinoma can have different outcomes at the same TNM stage. The authors developed TIGERS, a nine-feature score derived from routine H&E whole-slide images that measures tumour burden, glandular and nuclear shape, chromatin texture and tumour-stroma arrangement. It was trained in 644 treatment-naive cases and tested in six other cohorts, including a post-neoadjuvant cohort and a prospectively accrued cohort of 155, using one formula and a locked cutoff.

The cutoff stratified survival in the internal validation cohort and three of four retrospective test cohorts. In the remaining external cohort, 226 of 235 patients (96.2%) were called high risk, so the fixed threshold did not transfer. Adding TIGERS improved Harrell's C-index in all six retrospective cohorts by 0.015 to 0.047. In the prospective cohort, mortality was 47.5% in the high-risk and 22.7% in the low-risk group (adjusted HR 4.45, 95% CI 2.28 to 8.70). Scoring took a median of 13.5 minutes.

It reclassified 43.2% of patients relative to the stage-based definition. The authors state that prospective interventional studies are needed before it can guide treatment or surveillance. For pathologists this is a preview of image-based adjuncts, with scanner and cutoff calibration the practical issue.

• The score added to clinicopathological models in all six retrospective cohorts (C-index gain 0.015 to 0.047).
• In the prospective cohort the adjusted hazard ratio for high risk was 4.45.
• The locked cutoff failed in one external cohort, with 96.2% called high risk.
• Calibration and scanner differences are the main barrier to local use.
• Not for treatment decisions until interventional studies are done.

Why it matters

It suggests that slides pathologists already produce hold prognostic information beyond stage.

Don't overread it

The score has not been shown to change treatment, and failure to transfer in one cohort shows the cutoff is sensitive to the setting.

The statistics, in plain English

A C-index gain of 0.015 to 0.047 means the score improved the model's ability to rank who survives longer, but by a modest amount. The hazard ratio of 4.45 (95% CI 2.28 to 8.70) in 155 patients has a wide interval, reflecting the small prospective group.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.