A team from China used AI methods to retrospectively study cytology from 57,000 cases of benign or malignant effusions.

Cancer of unknown primary site often causes pleural and peritoneal effusions, and cytological analysis by pathologists using traditional methods can be inaccurate. Researchers in China used various deep- learning, artificial intelligence (AI) methods to retrospectively study cytology from 57,000 cases of be- nign or malignant effusions, in which the primary tumor initially was unknown but was determined subsequently. The ability of the AI system to identify both malignant and benign cells and tumor origin then was tested retrospectively on nearly 30,000 additional samples.

AI technology was highly accurate at identifying malignancies and tumor origins — as accurate as senior pathologists, and significantly more accurate than junior pathologists. The junior pathologists’ performance on new samples without receiving AI predictions was reevaluated after they were given the AI results and asked to reconsider their diagnoses — AI greatly improved diagnostic accuracy. More- over, routine use of AI likely would have improved outcomes: Patients with cancer of unknown primary site who received treatment concordant with tumor origins predicted by AI had significantly longer overall survival than those whose treatment was discordant with AI predictions (27 vs. 17 months).

COMMENT

AI proved remarkably accurate in identifying the presence of cancer and the tumor of origin in pleural or peritoneal effusions when initial AI results were compared with subsequent evidence of cancer and the cancer’s tissue of origin. Although the results are limited to the analysis of those two types of effu- sions and to the cancers that typically produce those effusions, we are likely to see more uses of AI in pathology laboratories. — Anthony L. Komaroff, MD

Dr. Komaroff is Professor of Medicine at Harvard Medical School, Boston.

Tian F et al. Prediction of tumor origin in cancers of unknown primary origin with cytology-based deep learning. Nat Med 2024 Apr 16; [e-pub]. (https://doi.org/10.1038/s41591-024-02915-w)

GKB-NON-2024-00447