Whole slide image-level classification of malignant effusion cytology using clustering-constrained attention multiple instance learning

Background: Cytological diagnosis of pleural effusion plays an important role in the early detection and diagnosis of lung cancers. Recently, attempts have been made to overcome low diagnostic accuracy and interobserver variability using artificial intelligence-based image analysis. However, such analysis is primarily performed at the image-patch level and not at the whole-slide image (WSI) level. This study aims to develop a WSI-level classification of malignant effusions in metastatic lung cancer based on pleural fluid cytology using a quality-controlled, nationwide dataset. Methods: The dataset was collected by a consortium research group that included three major university hospitals and the Quality Assurance Program Committee of the Korean Society of Cytopathology. It contains 576 normal and 309 cancer WSIs from pleural fluids. A clustering-constrained attention multiple-instance learning (CLAM) model was used for WSI-level classification. Results: The CLAM model achieved a high accuracy of 97%, with an area under the curve of 0.97, representing a 13% improvement over the image patch classification model-based WSI classification. It also significantly reduced the analysis time and computing resources compared to those required during image patch-level classification and heat map generation on the WSIs. Conclusion: The CLAM model successfully demonstrated high performance in differentiating malignant effusion at the WSI level using a large, quality-controlled, nationwide dataset. Further external validation is required to ensure generalizability.