Fine-needle aspiration (FNA) of the pancreas is the gold standard for diagnosing pancreatic ductal adenocarcinoma (PDAC), the most common form of pancreatic cancer. A pathologist analyzes the aspirated cells under a microscope, but the stakes are extremely high — a suspicion of malignancy can lead directly to chemotherapy and major surgery.
Digitally scanned cytopathology slides are enormous, reaching 50,000 to 100,000 pixels in each dimension. This gigapixel scale creates a major challenge for AI models, which must process the entire slide without missing suspicious areas.
The researchers developed a novel deep learning model called MIPCL (Multiple Instance Patch-level Classification Learning) that processes whole slide images at full resolution. MIPCL works by encoding tile-level features, separating foreground cancer cells from background tissue, and then pooling high-confidence cancer regions for a final slide-level prediction.
The team also devised a smart data curation strategy unique to cytopathology: since each FNA case produces multiple slides with distinct geographic regions of normal and abnormal cells, these regions were treated as independent training examples, significantly increasing the usable dataset without collecting additional patient cases.
MIPCL achieved an F1-Score of 91.07% and an AUROC of 0.9435, outperforming two established deep learning models (ABMIL and CLAM) which scored F1-Scores of 87.97% and 88.70% respectively. These gains were statistically significant.
Importantly, the model also uncovered 14 cases that had been incorrectly labeled as cancer-negative during data curation — suggesting the AI may be catching real cancers that human review missed. The model visualizes its most influential prediction regions, enabling pathologists to focus on the highest-yield areas of the slide.
There is a growing shortage of cytotechnologists — trained professionals who screen cytopathology slides before expert review. An AI triaging tool like MIPCL could help prioritize which cases need immediate attention, easing the burden on understaffed pathology labs.
The model's ability to highlight the most cancer-suspicious regions on a slide also serves as a teaching aid and quality-control tool, helping less-experienced staff improve their diagnostic accuracy.
MIPCL demonstrates that deep learning can be successfully applied to the demanding task of cytopathology classification at the whole-slide level, a frontier that prior studies had not fully addressed. The model works on both Diff-Quik and Pap smear stains.
Future work should validate the model across multiple institutions and explore whether the AI can support other cytopathology tasks beyond cancer versus non-cancer classification, such as predicting tumor subtype.