TY - JOUR
T1 - Stroma and lymphocytes identified by deep learning are independent predictors for survival in pancreatic cancer
AU - Tan, Xiuxiang
AU - Rosin, Mika
AU - Appinger, Simone
AU - Deierl, Julia Campello
AU - Reichel, Konrad
AU - Coolsen, Mariëlle
AU - Valkenburg-van Iersel, Liselot
AU - de Vos-Geelen, Judith
AU - de Jong, Evelien J.M.
AU - Bednarsch, Jan
AU - Grootkoerkamp, Bas
AU - Doukas, Michail
AU - van Eijck, Casper
AU - Luedde, Tom
AU - Dahl, Edgar
AU - Kather, Jakob Nikolas
AU - Sivakumar, Shivan
AU - Knoefel, Wolfram Trudo
AU - Wiltberger, Georg
AU - Neumann, Ulf Peter
AU - Heij, Lara R.
N1 - Publisher Copyright:
© The Author(s) 2025.
PY - 2025/3/19
Y1 - 2025/3/19
N2 - 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 disease for months or years. Thus, there is a clinical need to search corresponding prognostic biomarkers which forecast survival on an individual basis. To dig more information and identify potential biomarkers from PDAC pathological slides, we trained a deep learning (DL) model based U-net-shaped backbone. This DL model can automatically detect tumor, stroma and lymphocytes on whole slide images (WSIs) of PDAC patients. We performed an analysis of 800 PDAC scans, categorizing stroma in percentage (SIP) and lymphocytes in percentage (LIP) into two and three categories, respectively. The presented model achieved remarkable accuracy results with a total accuracy of 94.72%, a mean intersection of union rate of 78.66%, and a mean dice coefficient of 87.74%. Survival analysis revealed that SIP-mediate and LIP-high groups correlated with enhanced median overall survival (OS) across all cohorts. These findings underscore the potential of SIP and LIP as prognostic biomarkers for PDAC and highlight the utility of DL as a tool for PDAC biomarkers detecting on WSIs.
AB - 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 disease for months or years. Thus, there is a clinical need to search corresponding prognostic biomarkers which forecast survival on an individual basis. To dig more information and identify potential biomarkers from PDAC pathological slides, we trained a deep learning (DL) model based U-net-shaped backbone. This DL model can automatically detect tumor, stroma and lymphocytes on whole slide images (WSIs) of PDAC patients. We performed an analysis of 800 PDAC scans, categorizing stroma in percentage (SIP) and lymphocytes in percentage (LIP) into two and three categories, respectively. The presented model achieved remarkable accuracy results with a total accuracy of 94.72%, a mean intersection of union rate of 78.66%, and a mean dice coefficient of 87.74%. Survival analysis revealed that SIP-mediate and LIP-high groups correlated with enhanced median overall survival (OS) across all cohorts. These findings underscore the potential of SIP and LIP as prognostic biomarkers for PDAC and highlight the utility of DL as a tool for PDAC biomarkers detecting on WSIs.
UR - http://www.scopus.com/inward/record.url?scp=105000285410&partnerID=8YFLogxK
U2 - 10.1038/s41598-025-94362-x
DO - 10.1038/s41598-025-94362-x
M3 - Article
C2 - 40108402
AN - SCOPUS:105000285410
SN - 2045-2322
VL - 15
JO - Scientific Reports
JF - Scientific Reports
IS - 1
M1 - 9415
ER -
