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

Area of Science:

• Pathology
• Oncology
• Medical Imaging

Background:

• Cytological diagnosis of pleural effusion is vital for early lung cancer detection.
• Artificial intelligence (AI) aims to improve accuracy and reduce variability in cytological analysis.
• Current AI approaches often focus on image patches, not whole-slide images (WSIs).

Purpose of the Study:

• To develop a WSI-level classification model for malignant effusions in metastatic lung cancer.
• To utilize a quality-controlled, nationwide dataset of pleural fluid cytology.
• To enhance diagnostic accuracy beyond image-patch level analysis.

Main Methods:

• A nationwide dataset of 576 normal and 309 cancer WSIs from pleural fluids was compiled.
• A clustering-constrained attention multiple-instance learning (CLAM) model was employed for WSI classification.
• The CLAM model was trained and evaluated on the quality-controlled dataset.

Main Results:

• The CLAM model achieved 97% accuracy and an AUC of 0.97 for WSI-level classification.
• This represents a 13% improvement compared to image-patch classification methods.
• The CLAM model significantly reduced analysis time and computational resources.

Conclusions:

• The CLAM model demonstrates high performance for WSI-level differentiation of malignant pleural effusion.
• The study utilized a large, quality-controlled, nationwide dataset.
• External validation is recommended to confirm generalizability.