159P - Classical computer vision and modern deep-learning of pancreatic stroma histology features to diagnose cancer

Background

The stromal component constitutes as much as 90% of pancreatic cancer specimens, dynamically interacting with the tumor and adapting into a pro-survival environment. This poses a clinical challenge, as biopsies often miss cancer by only sampling stroma. By leveraging AI and image analysis, we aim to extract informative cues from stromal interactions for novel cancer biomarker identification. This approach offers the potential for enhanced diagnostic precision and a deeper understanding of pancreatic cancer biology.

Methods

Anonymized digital scans of pancreatic cancer and chronic pancreatitis were sourced from the Centre Hospitalier de l’Université de Montréal. QuPath 0.4.3 aided slide annotation, with subsequent TIF annotation export. Staining normalization was performed via the Mitkovetta technique in Python. Our process involved deep-learning stromal segmentation, prioritizing >95% stromal tiles using Ilastik. Feature extraction was executed utilizing computer vision techniques (Haralick features), alongside the pre-trained and class-trained ImageNet deep-neural network, VGG16.

Results

Our annotated, normalized, automated, and 95% stroma-probability method generated for the training cohort 9829 cancer and 1638 mass-forming pancreatitis tiles, and 10776 cancer and 1211 pancreatitis tiles for the testing set. The table highlights the performance of the classical computer vision approach (Haralicks features extraction in RGB). Furthermore, transferring the ImageNet VGG-16 pre-trained model to our dataset managed to predict the presence of adjacent cancer at 86.6% accuracy. Table: 159P

Conclusions

We demonstrate that normalized stromal tiles could predict the presence of cancer accurately just by their morphological features at HE staining. This highlights the importance of stroma for diagnostic purposes and can serve as the basis for future studies through multiplex imaging and spatial transcriptomics.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Vincent Quoc-Huy Trinh.

Funding

Fonds de Recherche Québec Santé.

Disclosure

All authors have declared no conflicts of interest.