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

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Conventional ascitic fluid cytology has low sensitivity for ovarian cancer detection.
• There is a need for more accurate and sensitive diagnostic methods.

Purpose of the Study:

• To develop and evaluate artificial intelligence (AI) models for detecting ovarian cancer in ascitic fluid cytology.
• To compare the performance of convolutional neural network (CNN) and clustering-constrained attention multiple-instance learning (CLAM) algorithms.

Main Methods:

• Developed patch image (PI)-based CNN models and CLAM algorithms using whole-slide images (WSIs) of ascitic fluid.
• Trained and validated six CNN algorithms and two CLAM algorithms (CLAM-SB, CLAM-MB).
• Evaluated model performance using accuracy and area under the curve (AUC) on internal and external datasets.

Main Results:

• ResNet50 (a CNN model) achieved an accuracy of 0.973 for PI-based detection.
• CLAM-SB demonstrated superior performance on external validation data with an AUC of 0.866, outperforming ResNet50 (AUC 0.804).
• AI and human interpretation showed complementary strengths, suggesting enhanced diagnostic accuracy with computer-aided diagnosis.

Conclusions:

• AI models, particularly CLAM, show significant potential to improve ovarian cancer detection sensitivity and accuracy in ascitic fluid cytology.
• WSI-based learning in CLAM eliminates the need for patch-by-patch annotation, offering an advantage over traditional CNN models.
• Computer-aided diagnosis is expected to enhance diagnostic accuracy and reproducibility in ovarian cancer detection.