Histologic classification of pancreatic cancer from cytology specimens requires expert pathologist review, which is time-consuming, subject to interobserver variability, and can be limited by specimen adequacy. This study introduces Slide2Graph, a graph convolutional neural network (GCN) framework designed to classify pancreatic cancer subtypes from whole-slide cytology images.
The study focused on specimens obtained by endoscopic ultrasound fine needle aspiration/biopsy (EUS-FNA/FNB) - the primary method for obtaining tissue from pancreatic masses. These cell block specimens present unique challenges for automated analysis due to sparse cellularity and variable fixation compared to surgical resection specimens.
The framework aims to distinguish pancreatic ductal adenocarcinoma (PDAC) from pancreatic neuroendocrine tumors (PanNET) and other entities, providing a tool that could assist pathologists in high-volume centers or enable diagnostic support in settings with limited specialized pathology expertise.
The Slide2Graph pipeline begins with patch-level feature extraction using a ResNet-18 convolutional neural network. Each whole-slide image is divided into non-overlapping patches (tiles), and ResNet-18 extracts a high-dimensional feature vector from each patch, capturing local cellular morphology without requiring manual annotation of individual cells.
These patch-level features are then used to construct a graph, where each node represents a patch and edges connect spatially adjacent patches. A graph convolutional network (GCN) then processes this graph structure, allowing the model to capture spatial relationships between tissue regions - a key advantage over bag-of-patches approaches that ignore spatial context.
The GCN aggregates information across the entire slide graph to produce a slide-level classification. This hierarchical approach - local features (ResNet-18) - spatial structure (graph) - whole-slide decision (GCN) - mirrors the way pathologists integrate local cellular details with tissue architecture patterns across the entire specimen to reach a diagnosis.
The study was conducted on 143 digitized cytology slides from EUS-FNA/FNB cell block specimens, representing a challenging and clinically realistic dataset. Cell blocks from EUS procedures contain compressed tissue fragments with variable quality, making them harder to analyze computationally than standard surgical pathology sections.
The dataset included cases of PDAC, PanNET, and other pancreatic lesions, with histopathologic diagnoses confirmed by expert pathologists serving as the ground truth. Digitization was performed using whole-slide imaging scanners, enabling high-resolution analysis at 20x or 40x equivalent magnification.
Data preprocessing included tissue detection (separating tissue from background glass), normalization to reduce staining variability between slides prepared at different times or institutions, and patch extraction at the appropriate resolution. These preprocessing steps are critical for ensuring the neural network focuses on diagnostically relevant tissue rather than slide artifacts.
The Slide2Graph model achieved an F1 score of 0.85 for histologic classification across the test set. The F1 score combines precision and recall into a single metric, making it particularly informative for imbalanced datasets where accuracy alone can be misleading. An F1 of 0.85 indicates a strong balance between sensitivity (detecting true positive cases) and specificity (avoiding false positives).
The model successfully classified both PDAC and PanNET from EUS-FNA/FNB cell block slides, demonstrating that graph-based analysis can capture the distinctive architectural and cellular features that differentiate these histologically distinct tumor types despite the challenging nature of cytology specimens.
Compared to baseline methods that treated whole-slide images as simple bags of patches (ignoring spatial structure), Slide2Graph's graph-based approach showed improved performance, validating the hypothesis that spatial tissue organization carries important diagnostic information beyond individual cell-level features.
A key conceptual contribution of this work is demonstrating that graph-based representations of histology slides can capture diagnostically important tissue architecture patterns that are lost when slides are analyzed as unordered collections of patches. In pancreatic pathology, the spatial arrangement of ductal structures, desmoplastic stroma, and cellular clusters carries significant diagnostic meaning.
The graph construction strategy - connecting spatially adjacent patches - effectively encodes 2D spatial relationships into the graph topology. The GCN then propagates information between neighboring nodes, allowing the model to recognize patterns that span multiple adjacent tissue regions, analogous to how pathologists scan across a slide at low magnification before zooming in.
Future extensions could incorporate multi-scale graph hierarchies (combining patch-level and region-level graphs), incorporate molecular features from immunohistochemistry, or use self-supervised pre-training on large unlabeled pathology datasets to improve feature representations before fine-tuning on the classification task.
In clinical practice, Slide2Graph could serve as a second-reader tool to assist pathologists reviewing EUS-FNA/FNB cell block specimens. Given the high stakes of pancreatic mass diagnosis - where misclassification can lead to incorrect surgical planning or delayed treatment - even modest improvements in diagnostic consistency have significant patient impact.
The system could be particularly valuable in lower-volume centers or settings where subspecialty pancreatic pathology expertise is limited. Real-time classification of digitized cytology slides could provide a preliminary assessment within minutes of scan submission, flagging cases for priority review or providing confidence scores to assist pathologists in borderline cases.
Regulatory and workflow integration challenges remain significant. Adoption of AI pathology tools requires prospective clinical validation, integration with laboratory information systems, validation across different scanners and staining protocols, and clear guidelines for how AI outputs should influence final diagnostic decisions. Despite these challenges, graph neural network approaches represent a promising direction for computational pathology in pancreatic cancer.