Automatic quantification of tumor-stroma ratio as a prognostic marker for pancreatic cancer

Area of Science:

• Computational pathology
• Oncology
• Digital pathology

Background:

• Pancreatic cancer lacks robust prognostic biomarkers, necessitating new tools for patient outcome prediction.
• Current staging systems have limitations in accurately predicting patient survival.

Purpose of the Study:

• To develop and validate an automated pipeline for quantifying the tumor-stroma ratio (TSR) in pancreatic cancer.
• To assess the potential of TSR as an independent prognostic marker for pancreatic cancer survival.

Main Methods:

• A deep learning approach was used for automatic segmentation of tumor and stroma components from whole-slide images.
• The pipeline was trained and validated using five-fold cross-validation and an independent external test set.
• Tumor-stroma ratio (TSR) was calculated, and its association with six-month survival was evaluated.

Main Results:

• The deep learning models achieved high segmentation accuracy, with median Dice scores of 0.751 and 0.726 for tumor epithelium and 0.76 and 0.863 for tumor bulk on internal and external test sets, respectively.
• The automated TSR quantification demonstrated potential as an independent prognostic marker.
• Cross-validation showed an AUC of 0.61±0.12 for predicting six-month survival on the external dataset.

Conclusions:

• The automated TSR quantification pipeline shows promise as a prognostic biomarker for pancreatic cancer.
• Computational biomarker discovery can enhance patient outcome prediction and support personalized treatment strategies.