This study introduces TORCH, a deep-learning method for predicting tumor origin in cancers of unknown primary (CUP) site, utilizing cytological images from pleural and peritoneal effusions. Evaluated on internal and external testing sets, TORCH demonstrated high AUROC values (0.953–0.991 for cancer diagnosis and 0.953–0.979 for tumor origin localization) and accuracy (82.6% top-1, 98.9% top-3). TORCH outperformed pathologists, significantly enhancing junior pathologists' diagnostic scores. Concordant treatment with TORCH predictions correlated with improved overall survival in CUP patients. While promising, further validation is needed.

Fei Tian, Dong Liu, Na Wei, Qianqian Fu, Lin Sun, Wei Liu, Xiaolong Sui, Kathryn Tian, Genevieve Nemeth, Jingyu Feng, Jingjing Xu, Lin Xiao, Junya Han, Jingjie Fu, Yinhua Shi, Yichen Yang, Jia Liu, Chunhong Hu, Bin Feng, Yan Sun, Yunjun Wang, Guohua Yu, Dalu Kong, Meiyun Wang, Wencai Li, Kexin Chen, Xiangchun Li