Peritoneal washing cytology (CY) involves flushing the abdominal cavity with fluid and examining the cells recovered under a microscope. Finding cancer cells in this fluid (CY-positive) indicates that cancer has spread into the peritoneal space and is classified as distant metastasis — these patients have a particularly poor prognosis.
Even among CY-positive patients, however, survival varies considerably. Some patients survive well beyond one year while others do not. This study asked whether deep learning analysis of CY specimen images — looking at all the cells present, not just the cancer cells — could predict which CY-positive patients were most likely to survive beyond one year.
88 CY-positive pancreatic cancer patients from a single Japanese center were enrolled, divided into poor prognosis (survival under 1 year; n=28) and good prognosis (survival 1 year or more; n=60) groups. CY specimens were stained, scanned at high magnification (40x) using a whole-slide imaging device, and then individual cell nuclei were computationally segmented.
Two deep learning architectures were trained and compared: a Vision Transformer (ViT), which uses an attention mechanism to analyze image patches, and a standard Convolutional Neural Network (CNN). Both were trained on 59,182 cell nuclei from the training cohort and evaluated on 17,084 nuclei from the test cohort.
Rather than training the AI on the whole slide, each individual cell nucleus was classified as likely poor-prognosis or good-prognosis. If enough nuclei from a given patient's specimen were classified as poor-prognosis, the overall specimen was judged to predict poor survival — linking single-cell analysis to patient-level outcomes.
Both models predicted 1-year prognosis with AUC above 0.80: ViT achieved 0.806 and CNN achieved 0.801. On an independent validation cohort from 2022, ViT's AUC was 0.833. Kaplan-Meier survival analysis confirmed that patients classified by ViT as good prognosis lived significantly longer than those classified as poor prognosis (p<0.0001).
To understand what the AI had learned, the researchers extracted the five cell nuclei most strongly associated with poor and good prognosis predictions. Cells associated with poor prognosis showed segmented, multi-lobed nuclei — the hallmark of neutrophils. This was confirmed by expert pathologists examining the images.
Cells associated with good prognosis were round-nucleated and showed no expression of calretinin — a marker for mesothelial cells — confirming they were macrophages. This means the AI learned that the presence of neutrophils is a bad prognostic sign and macrophages are a good sign, both findings that are biologically meaningful.
The discovery that intraperitoneal neutrophils are associated with poor prognosis aligns with research showing that neutrophils can promote tumor growth by creating structures (neutrophil extracellular traps) that capture circulating tumor cells and help them implant in new locations.
This suggests that targeting intraperitoneal neutrophils — for example, by enhancing their anti-tumor functions through type 1 interferon or by blocking their tumor-promoting activities — could be a new therapeutic strategy specifically for CY-positive pancreatic cancer patients. Therapies targeting tumor-associated neutrophils are currently being investigated.
The finding that macrophages in the peritoneal fluid associate with better outcomes suggests these cells may have an M1-like, tumor-fighting polarization state, which could potentially be enhanced therapeutically to improve survival.
This is the first study to demonstrate that deep learning of peritoneal washing cytology specimens can predict prognosis in pancreatic cancer, with accuracy comparable to established AI approaches for other cancer types. The approach adds new clinical value to a specimen type that is routinely collected during staging laparoscopy.
Future work will need to validate these findings in larger, multi-center cohorts and explore whether combining CY image analysis with genomic or clinical data can further improve prediction accuracy. If validated, this approach could guide treatment decisions — directing patients predicted to have poor prognosis toward more aggressive early interventions.