Deep Learning Predicts Tumor Origin and Malignancy in Cytology and Histopathology WSIs
Overview
The MAMILE-UNI deep learning framework demonstrated high accuracy and robust performance in detecting malignancy and predicting tumor origin from pleural and ascitic cytology and histopathology whole slide images (WSIs). Evaluations on over 2,400 WSIs showed significant diagnostic improvements validated by statistical testing.
Background
Pleural and ascitic cytology are critical for diagnosing metastatic cancers and determining tumor origin, but traditional microscopic examination suffers from low accuracy and observer variability. Deep learning approaches have shown promise in pathology but remain underutilized in effusion cytology. Accurate identification of malignancy and tumor origin from cytology smears, cell blocks, and histopathological slides can improve diagnostic workflows and patient management.
Data Highlights
Evaluation Dataset
Number of WSIs
Metrics
Performance
Malignancy detection in cytology (smear/cell block)
1250
AUROC, Mean Sensitivity & Specificity (MeanSS), Accuracy
High (exact values not specified)
Cancer origin identification from cytology smears
Not specified
Accuracy, MeanSS, AUROC
High
Cancer origin identification from histopathology WSIs
1196
Accuracy, Precision, Sensitivity, F1 Score, Specificity, MeanSS, AUROC
High; predictions validated by Fisher’s exact test (p < 0.001)
Key Findings
MAMILE-UNI achieved high diagnostic accuracy in detecting malignancy directly from pleural and ascitic cytology WSIs.
The model accurately identified the tumor origin from cytology smear images, demonstrating strong classification performance.
In histopathological WSIs, the framework maintained high precision, sensitivity, specificity, and overall accuracy for tumor origin prediction.
Statistical validation using Fisher’s exact test confirmed the significance of model predictions (p < 0.001).
The approach is data-efficient and implemented using widely available open-source tools and libraries.
Public availability of the program code supports reproducibility and further research.
Clinical Implications
The MAMILE-UNI deep learning framework can enhance diagnostic accuracy and reduce observer variability in evaluating pleural and ascitic effusions. Its ability to predict tumor origin from both cytology and histopathology WSIs may streamline diagnostic workflows and support personalized treatment planning. Integration of such AI tools could improve early detection and management of metastatic cancers.
Conclusion
MAMILE-UNI represents a promising AI-driven approach for malignancy detection and tumor origin prediction in cytology and histopathology WSIs, offering high accuracy and validated clinical relevance. This technology has potential to augment traditional pathology and improve patient outcomes.
References
Roberts et al. 2010 -- Management of a malignant pleural effusion: British thoracic society pleural disease guideline
Oey et al. 2016 -- The diagnostic work-up in patients with ascites: current guidelines and future prospects
Kassirian et al. 2023 -- Diagnostic sensitivity of pleural fluid cytology in malignant pleural effusions: systematic review and meta-analysis
Xie et al. 2022 -- Deep convolutional neural network-based classification of cancer cells on cytological pleural effusion images
