Deep Learning Enabled Scoring of Pancreatic Neuroendocrine Tumors Based on Cancer Infiltration Patterns

Endocrine Pathology 2025 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Page [1, 2]
Reading How Cancer Invades Tissue — A New AI Pathology Approach

When pathologists examine cancer tissue under the microscope, they assess not just the cancer cells themselves but how they invade the surrounding tissue. The pattern of tumor infiltration — whether cancer grows in smooth, pushing fronts or sends irregular, finger-like projections into surrounding tissue — correlates with how aggressively the cancer behaves and how likely it is to metastasize or recur.

Pancreatic neuroendocrine tumors (PanNETs) are graded primarily by cell proliferation rates (Ki-67 index and mitotic count) under WHO criteria. However, researchers have proposed that infiltration pattern scoring offers additional prognostic information beyond what these proliferation markers capture. Tumors with infiltrative growth patterns recur more often even when their proliferation grade appears low.

This study developed a deep learning system to automatically score PanNET infiltration patterns from digitized pathology slides — a task currently done subjectively by expert pathologists — aiming to make this additional prognostic measure objective, reproducible, and scalable.

TL;DR: Tumor infiltration patterns in PanNETs carry prognostic information beyond standard grading; deep learning was developed to automatically and objectively score these patterns from pathology slides.
Pages 3-3
Teaching AI to See How Cancer Invades Tissue on Digital Pathology Slides

Whole slide images (WSIs) of PanNET surgical specimens were digitized at high resolution, creating the dataset for AI training. Expert pathologists first manually annotated infiltration patterns — labeling regions of tumor margin as either expansile (smooth, pushing) or infiltrative (irregular, finger-like) — providing ground truth labels for the deep learning model.

The deep learning pipeline processed the large WSIs by dividing them into smaller tiles (patches) that could be analyzed individually by a convolutional neural network. The network was trained to classify each tile according to its infiltration pattern, then tile-level predictions were aggregated to produce an infiltration score for the whole tumor.

The resulting score was then correlated with patient outcomes — disease-free survival and overall survival — and compared to the standard WHO grading system to assess whether infiltration pattern scoring provides independent or complementary prognostic information.

TL;DR: Pathologist-annotated whole slide images trained a convolutional neural network to classify tumor infiltration patterns, which were then validated against patient survival outcomes.
Pages 5-5
AI-Scored Infiltration Patterns Predict Recurrence Beyond Standard Grading

The deep learning infiltration score successfully stratified PanNET patients into distinct risk groups with significantly different recurrence rates and overall survival. Tumors classified as having predominantly infiltrative patterns had worse outcomes, consistent with the biological expectation that irregular invasion reflects more aggressive behavior.

Importantly, the infiltration score provided prognostic information independent of WHO grade — meaning that even within the same grade category, patients with high infiltration scores had worse outcomes than those with lower scores. This suggests the two measures capture different aspects of tumor biology that are both clinically relevant.

Interobserver agreement among pathologists for manual infiltration scoring was imperfect, highlighting the subjective nature of this assessment. The deep learning model, by contrast, produces consistent scores every time — addressing a fundamental limitation of human pathological grading.

TL;DR: AI infiltration scoring predicted PanNET recurrence independently of WHO grade, and achieved better consistency than pathologist-assigned infiltration assessments.
Page [3, 4]
Processing Gigapixel Pathology Images With Deep Learning Patch Analysis

Digital whole slide pathology images are enormous — a single slide can be a gigapixel image (1,000+ megapixels) that cannot be processed by a standard neural network designed for regular photographs. The patch-based approach divides each slide into thousands of small, overlapping tiles that are each independently classified by the network.

Multiple-instance learning (MIL) techniques were applied to aggregate tile-level predictions without requiring every tile to be individually annotated. The model learns which patches within a slide are diagnostically informative and weights them accordingly when computing the final slide-level infiltration score.

Transfer learning from models pretrained on large natural image datasets (ImageNet) and pathology-specific pretraining (using other pathology image collections) provided the foundational visual feature representations, which were then fine-tuned on the PanNET infiltration classification task.

TL;DR: Patch-based processing with multiple-instance learning allowed deep learning to analyze gigapixel pathology slides and produce consistent whole-tumor infiltration scores.
Page [6, 7]
Making Pathological Grading More Objective and Reproducible

Current pathological grading of PanNETs involves subjective judgments that vary between pathologists, particularly for borderline cases. An AI system that consistently quantifies infiltration patterns could reduce this variability, ensuring that patients receive consistent prognostic assessments regardless of which pathologist reviews their slides.

In practice, the deep learning infiltration score could be reported alongside the standard WHO grade as a complementary prognostic metric — similar to how Oncotype DX genomic scores are reported alongside conventional breast cancer staging. This additional information could influence decisions about surveillance frequency, adjuvant therapy, or participation in clinical trials.

The system could also be applied retrospectively to archived slides from clinical trials, enabling investigators to analyze whether infiltration pattern predicted response to specific treatments — potentially identifying a new biomarker for treatment selection in PanNET patients.

TL;DR: The AI infiltration score could complement standard WHO grading as a consistent, objective metric that improves prognostic precision for PanNET patients.
Page [7, 8]
AI Pathology Adds a New Dimension to Pancreatic Neuroendocrine Tumor Assessment

This study demonstrates that deep learning can automate the scoring of PanNET infiltration patterns from digital pathology slides, generating consistent and prognostically meaningful assessments that complement standard WHO grading. The AI score captures tumor biological behavior that human pathologists acknowledge as important but assess inconsistently.

The combination of standard WHO grading and AI infiltration scoring provides a richer characterization of PanNET biology than either alone, potentially identifying patients at higher recurrence risk who might benefit from more intensive surveillance or adjuvant treatment even when their tumors appear to be low grade by conventional criteria.

With the continued digitization of pathology archives and the increasing adoption of digital pathology platforms in clinical practice, deploying AI tools like this infiltration scoring system requires only software integration rather than new equipment — making clinical translation both feasible and timely.

TL;DR: Deep learning-based infiltration scoring adds prognostic value beyond standard PanNET grading and is ready for clinical translation as pathology workflows increasingly go digital.
Citation: Open Access, 2025. Available at: PMC11757657.