Endoscopic Ultrasound-guided Fine Needle Aspiration (EUS-FNA) is the standard procedure for obtaining tissue samples from pancreatic masses. During the procedure, a thin needle is guided by ultrasound through the stomach wall to sample the pancreatic lesion in a minimally invasive manner.
Rapid On-Site Cytopathology Evaluation (ROSE) involves immediately staining and examining the needle aspirate under a microscope during the biopsy procedure. ROSE allows cytopathologists to confirm that adequate diagnostic material has been collected, reducing the need for repeat procedures.
A major limitation of ROSE is that it requires a trained cytopathologist to be physically present in the endoscopy suite during the procedure. This expertise is not available at all hospitals, creating disparities in the quality of pancreatic cancer diagnosis across different healthcare settings.
A deep learning system capable of performing automated ROSE analysis could extend high-quality pancreatic cancer diagnosis to hospitals without on-site cytopathology expertise, potentially improving diagnostic accuracy and reducing procedural time at a broad range of centers.
The study assembled a large multicenter dataset of 5,345 cytology slides from 194 patients across multiple hospitals. This multicenter design is important for evaluating whether deep learning models generalize across different staining protocols and slide preparation practices.
The deep learning system uses a UNet architecture with a ResNet-101 encoder backbone. UNet is an encoder-decoder network originally designed for biomedical image segmentation, while ResNet-101 is a powerful feature extractor pre-trained on millions of natural images and adapted for the cytology domain through fine-tuning.
The model performs a two-stage task: first, it segments individual cells within each cytology slide, delineating cell boundaries to isolate individual cellular objects. Then it classifies each slide based on aggregated cellular features to determine whether the sample contains malignant cancer cells or benign material.
For the cell segmentation task, the DCNN achieved an F1-score of 0.929, indicating that the model could accurately identify and delineate individual cells within cytology slides with high precision and recall. Accurate cell segmentation is foundational because subsequent cancer classification depends on correctly extracted cellular features.
For the cancer identification task, the model achieved an Area Under the Curve (AUC) of 0.958, demonstrating highly effective discrimination between malignant and benign pancreatic cytology specimens. This level of performance is clinically meaningful for a screening or triage application.
Performance was consistent across samples from different participating institutions, suggesting that the model learned generalizable features rather than institution-specific artifacts. This cross-site generalizability is a key requirement for real-world deployment of AI diagnostic tools.
The study directly compared the DCNN against two categories of human examiners: trained cytopathologists with specialized expertise in interpreting pancreatic cytology, and experienced endoscopists who perform EUS-FNA procedures regularly but lack specialized cytology training.
The DCNN performed comparably to cytopathologists on cancer identification, with overlapping confidence intervals suggesting that AI-level performance matches specialist human performance. This finding is significant because it suggests the system could serve as a credible expert stand-in at facilities without on-site cytopathology coverage.
The DCNN substantially outperformed trained endoscopists, who showed considerably lower accuracy when interpreting cytology slides. This comparison highlights the specificity of cytopathology expertise and supports the clinical value of automating ROSE analysis to exceed the capability of non-specialist clinicians.
Deploying an AI ROSE system in clinical practice could enable same-session diagnosis of pancreatic masses at hospitals that currently cannot offer on-site cytopathology support. Patients at such centers would benefit from faster results and fewer repeat biopsy procedures.
The model processes slide images in near real-time, making it technically feasible for intraoperative use during the EUS-FNA procedure. Integration with digital pathology scanners would allow the AI to analyze slides as they are prepared, providing diagnostic feedback within minutes.
A remaining challenge is the rare occurrence of ambiguous or indeterminate cytology specimens that are difficult even for expert cytopathologists. The system should be designed with an explicit uncertainty output that flags difficult cases for human review rather than forcing a binary malignant or benign classification.
Regulatory approval, physician acceptance, and appropriate workflows for human oversight are all necessary for clinical translation. The AI system is best positioned as a decision support tool that enhances rather than replaces cytopathologist judgment, particularly in resource-limited settings.