Ovarian Cancer Detection in Ascites Cytology with Weakly Supervised Model on Nationwide Data Set

Abstract

Conventional ascitic fluid cytology for detecting ovarian cancer is limited by its low sensitivity. To address this issue, this multicenter study developed patch image (PI)-based fully supervised convolutional neural network (CNN) models and clustering-constrained attention multiple-instance learning (CLAM) algorithms for detecting ovarian cancer using ascitic fluid cytology. Whole-slide images (WSIs), 356 benign and 147 cancer, were collected, from which 14,699 benign and 8025 cancer PIs were extracted. Additionally, 131 WSIs (44 benign and 87 cancer) were used for external validation. Six CNN algorithms were developed for cancer detection using PIs. Subsequently, two CLAM algorithms, single branch (CLAM-SB) and multiple branch (CLAM-MB), were developed. ResNet50 demonstrated the best performance, achieving an accuracy of 0.973. The performance when interpreting internal WSIs was an area under the curve (AUC) of 0.982. CLAM-SB outperformed CLAM-MB with an AUC of 0.944 for internal WSIs. Notably, in the external test, CLAM-SB exhibited superior performance with an AUC of 0.866 compared with ResNet50's AUC of 0.804. Analysis of the heatmap revealed that cases frequently misinterpreted by AI were easily interpreted by humans, and vice versa. Because AI and humans were found to function complementarily, implementing computer-aided diagnosis is expected to significantly enhance diagnostic accuracy and reproducibility. Furthermore, the WSI-based learning in CLAM, eliminating the need for patch-by-patch annotation, offers an advantage over the CNN model.