Stroma and lymphocytes identified by deep learning are independent predictors for survival in pancreatic cancer
Pancreatic ductal adenocarcinoma (PDAC) is one of the most lethal cancers known to humans. However, not all patients fare equally poor survival, and a minority of patients even survives advanced disea
Summary
Studies disagree on whether the desmoplastic stroma of pancreatic ductal adenocarcinoma (PDAC) is good or bad for the patient. Most were small and relied on manual estimates of the tumour–stroma ratio.
This multicentre study trains U-Net models to label tumour, stroma and lymphocytes on H&E whole-slide images, then applies them to 800 resected PDACs from four cohorts: Aachen, Düsseldorf, Rotterdam and TCGA. Two readouts, stroma in percentage (SIP) and lymphocytes in percentage (LIP), separated overall survival in every cohort:
Stroma: the relationship was U-shaped. An intermediate stroma proportion did best; both low and high did worse.
Lymphocytes: a high proportion was favourable.
In multivariable Cox models, SIP was an independent predictor in all four cohorts and LIP in three.
The abstract in the
description:frontmatter is the paper's own, verbatim. The sections below are my own-words digest of the abstract and the open-access full text in PubMed Central.
Citation
Tan X, Rosin M, Appinger S, Campello Deierl J, Reichel K, Coolsen M, et al. Stroma and lymphocytes identified by deep learning are independent predictors for survival in pancreatic cancer. Sci Rep. 2025;15(1):9415. Published 2025 Mar 19. doi: 10.1038/s41598-025-94362-x. PMID: 40108402. PMCID: PMC11923104.
Access: open access, full text in PubMed Central
Study at a glance
Slides
800 H&E slides of resected PDAC: Aachen (244), Düsseldorf (200), Erasmus MC Rotterdam (192), TCGA (164)
Region of interest
Senior pathologists outlined an area containing tumour glands, lymphocytes and desmoplastic stroma in QuPath
Training data
10 randomly selected WSIs per cohort, annotated for stroma, lymphocytes and tumour; more than 50,000 tiles per cohort (99 × 99 µm, 396 px at 0.25 µm/px); Macenko colour normalisation
Model
Modified U-Net (Keras), one model trained per cohort
Cut-offs
SIP: X-tile into low / intermediate / high, then low and high merged. LIP: maximally selected rank statistics into low / high
Key findings
Segmentation (TCGA test set): accuracy 94.72%, mean IoU 78.66%, mean Dice 87.74%.
Intermediate stroma range. Roughly 53–76% stroma, depending on the cohort.
Aachen
24 vs 12
22 vs 13
28
Düsseldorf
24 vs 9
24 vs 15
31
Rotterdam
27 vs 16
27 vs 16
33
TCGA
24 vs 13
23 vs 13
35
Multivariable Cox. SIP was independent in all four cohorts; LIP in Aachen, Rotterdam and TCGA.
A possible explanation for the U-shape. An exploratory deconvolution of TCGA bulk RNA-seq linked more inflammatory CAFs to better survival and more myofibroblastic CAFs to worse. The authors offer this stromal heterogeneity as a reason both extremes of stroma do badly.
Abstract vs Methods. The abstract says SIP and LIP were split into "two and three categories, respectively"; the Methods describe the reverse (SIP into three groups, later merged to two; LIP into two).
Limitations noted by the authors
Performance may vary with image quality and slide preparation between institutions.
Only three tissue classes were modelled; other cell populations were not identified.
The study shows survival associations only, with no functional or multimodal data.
Reading notes
My own notes, not content taken from the paper.
The cut-offs were chosen within each cohort to maximise survival separation, and each cohort had its own model. The thresholds are data-derived and would need testing on a fixed, pre-specified cut-off.
On the lecture slide
From a lecture slide titled "Top quantitative models in GI/liver path":
Pancreatic ductal adenocarcinoma
Tan U-Net
Stroma %, lymphocyte %
10 WSIs/cohort
Tan et al. Sci Rep. 2025
"10 WSIs/cohort" matches the annotated training slides; the models were then applied to all 800 slides.
Last updated