Precise Detection of Peritoneal Metastasis of Gastric and Colorectal Cancer Using Artificial Intelligence of Cytopathologic Whole Slide Images

Résumé fourni par la source

Background: Peritoneal metastasis is a common manifestation of advanced gastric cancer (GC) and colorectal cancer (CRC).Radiologic detection is often inconclusive, particularly with minimal ascites or subtle peritoneal thickening.Although ascitic fluid cytology is the diagnostic gold standard, it has low sensitivity, high observer variability, and depends on manual review.Artificial intelligence (AI) shows a promise for improving diagnosis, but patch-based models require segmentation and reconstruction, increasing computational load and reducing efficiency particularly with few malignant cells.Methods: We developed a whole slide image (WSI)-level diagnostic model for detecting metastatic GC and CRC in ascitic cytology using the Clustering-constrained Attention Multiple Instance Learning.This architecture identifies key regions without patch-wise annotation or reconstruction.We trained and validated it on the Open AI Dataset Project, a national cytology dataset comprising 243 malignant and 476 negative WSIs, divided into training, validation, and test sets.Results: The model achieved accuracies of 1.00 (validation) and 0.933 (test), with area under the receiver operating characteristic curves of 1.00 and 0.925.WSI-level MIL approaches may conceptually reduce computational burden by avoiding explicit patch-level classification.Compared to five experienced pathologists, the model achieved superior or comparable accuracy, and AI assistance improved pathologists' average diagnostic accuracy from 76.5% to 91.5%, while reducing interpretation time by over 40%.No false negatives occurred, and attention maps confirmed the malignant areas.Conclusion: Our WSI-level AI model offers a robust, efficient solution for diagnosing metastatic GC and CRC in ascitic cytology.It addresses limitations of manual and patchbased methods, supports clinical decision-making, and enhances diagnostic accuracy.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.