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

Abstract

Purpose: This study aims to introduce an innovative multi-step pipeline for automatic tumor-stroma ratio (TSR) quantification as a potential prognostic marker for pancreatic cancer, addressing the limitations of existing staging systems and the lack of commonly used prognostic biomarkers.

Methods: The proposed approach involves a deep-learning-based method for the automatic segmentation of tumor epithelial cells, tumor bulk, and stroma from whole-slide images (WSIs). Models were trained using five-fold cross-validation and evaluated on an independent external test set. TSR was computed based on the segmented components. Additionally, TSR's predictive value for six-month survival on the independent external dataset was assessed.

Results: Median Dice (inter-quartile range (IQR)) of 0.751(0.15) and 0.726(0.25) for tumor epithelium segmentation on internal and external test sets, respectively. Median Dice of 0.76(0.11) and 0.863(0.17) for tumor bulk segmentation on internal and external test sets, respectively. TSR was evaluated as an independent prognostic marker, demonstrating a cross-validation AUC of 0.61±0.12 for predicting six-month survival on the external dataset.

Conclusion: Our pipeline for automatic TSR quantification offers promising potential as a prognostic marker for pancreatic cancer. The results underscore the feasibility of computational biomarker discovery in enhancing patient outcome prediction, thus contributing to personalized patient management.

Fig 1. Example of dataset A (first row H&E stained WSI, corresponding CK18/8 stained WSI on the second row).

The staining-restaining procedure results in perfectly matching slides.

Fig 2

Flowchart highlighting different pipeline steps: (a) Epithelium segmentation and (b) tumor epithelium segmentation. Through the process of staining-destaining of paired H&E and IHC slides, epithelium annotations are obtained, which are then used to train an epithelium segmentation network (dataset A). This network annotates the rest of the slides. Subsequently, a tumor epithelium segmentation network is trained on the segmented epithelium combined with annotated tumor area (dataset B). Based on tumor epithelium segmentation, the tumor bulk is automatically determined by drawing a convex hull (c), on which TSR is calculated. Legend: In green the epithelium segmentation network, in yellow the tumor epithelium segmentation network and in red the resulting tumor bulk segmentation and TSR quantification.

Fig 3. Example of epithelial segmentation network on two different slides from dataset A.

Left column: example of two patches extracted from two different cases. Middle column: in green the overlay of the epithelium extracted from the CK18/8 stain. Right column: output of the network (in green the segmented epithelium). As we can see, lymphoid aggregates are correctly recognised as non-epithelium from the network.

Fig 4. Example of tumor epithelial segmentation network on a slide from dataset B.

First column: a randomly extracted patch from the dataset. Middle column: ground truth obtained by inferring the Epithelium segmentation network (top left cluster of epithelial cells are non cancerous, bottom right are cancerous). Right column: output of the tumor epithelium network. As we see, tumor epithelium is correctly segmented while healthy epithelium is discarded.

Fig 5. Example of the performance of the alphahull algorithm on dataset D.

Middle column is the coarse tumor annotation made by pathologist, while right column is the convex hull automatically generated on the segmented tumor epithelium.

Fig 6. AUC of the Logistic Regression model across five-fold cross-validation.

Shade in the area represents the standard deviation across all folds. The figure reports three experiments, 6 months, 12 months, and 18 months survival, respectively.

Fig 7. Kaplan-Meier estimator.

Stratification was done using the mean TSR value.