Deep learning for malignancy and tumor origin prediction using cytology or histopathology whole slide images

Pleural and ascitic cytology is essential for diagnosing metastatic cancer and predicting tumor origin, yet microscopic observation alone often leads to low accuracy and observer variability. Although deep learning shows great potential in pathology, its use in pleural and ascitic cytology remains limited. We present a data-efficient deep learning framework (MAMILE-UNI) that directly detects malignancy in pleural and ascitic effusions from cytology smear or cell block whole slide images (WSIs); in evaluation of 1250 WSIs, MAMILE-UNI achieved high AUROC, the mean of sensitivity and specificity (MeanSS), and accuracy. Furthermore, in identifying the origin of cancer from cytology smears, the method also achieved high accuracy, MeanSS and AUROC. Identifying the origin of cancer from histopathological slide images is equally important, and our method achieved high accuracy, precision, sensitivity, F1 score, specificity, MeanSS and AUROC in evaluation with 1,196 WSIs. Fisher's exact test validated the model predictions (p < 0.001).