A Robust Framework for Domain-Generalized Classification of Ovarian Cancer Histology Images.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in histology

Background:

• Computational pathology (CP) analysis demands high computational efficiency and accurate classification for scalable solutions.
• Whole-Slide Images (WSIs) present heterogeneity challenges, hindering robust generalization in deep learning models.
• Existing deep learning frameworks struggle with domain variability in WSIs.

Purpose of the Study:

• To present WSI-P2P (Whole-Slide Imaging-Patch to Prediction), a novel framework addressing computational efficiency and generalization in WSI analysis.
• To introduce the K-TOP MIL aggregator for selective processing of informative instances.
• To develop an adaptive feature extractor for handling multi-center dataset variability.

Main Methods:

• Leverages downscaled patch sampling and Multiple-Instance Learning (MIL) with transfer learning.
• Introduces the K-TOP MIL aggregator, a variant of attention-based MIL that processes the top K most informative instances.
• Features an online, adaptive feature extractor for end-to-end fine-tuning of pre-trained models.

Main Results:

• WSI-P2P achieves state-of-the-art accuracy, demonstrating superior domain adaptability and computational efficiency.
• The K-TOP MIL aggregator improved performance to approximately 100% AUROC and 95.72% test accuracy.
• The K-TOP MIL aggregator showed 2.3x computational efficiency compared to base aggregators, with consistent performance across domains.

Conclusions:

• WSI-P2P outperforms traditional offline feature extraction, maintaining high discriminative ability across diverse data distributions.
• The framework shows excellent performance in subtype classifications, proving its reliability for large-scale histological studies.
• WSI-P2P offers a scalable and reliable tool for clinical settings and extensive histological research.