Related Experiment Video
Updated: May 28, 2026
05:33
Introduction of an Integrated Pathology Image Management, Artificial Intelligence, and Reporting System
Published on: July 11, 2025
Comparative Evaluation of Feature Extractors, Aggregation Strategies, and Classification Hierarchies for Ovarian
Ho Jung Song1, You Sang Cho1, Yong Suk Kim2
1Department of Medical Engineering, Konyang University, 158 Gwanjeo-dong-ro, Seo-gu, Daejeon 32992, Republic of Korea.
Diagnostics (Basel, Switzerland)
|May 27, 2026
Summary
Feature extractor choice significantly impacts epithelial ovarian cancer subtype classification accuracy in whole slide images. Hierarchical classification frameworks consistently improve high-grade serous carcinoma detection rates.
Area of Science:
- Computational pathology
- Digital pathology
- Machine learning for medical imaging
Background:
- Multiple instance learning (MIL) is crucial for classifying epithelial ovarian cancer subtypes from whole slide images (WSIs).
- The impact of feature extractor, aggregation strategy, and classification framework on MIL performance, especially with imbalanced data, is not well understood.
Purpose of the Study:
- To evaluate the relative contributions of different components in multiple instance learning (MIL) pipelines for epithelial ovarian cancer (EOC) subtype classification.
- To compare the performance of various feature extractors, aggregation strategies, and classification frameworks under class imbalance.
Main Methods:
- Evaluated 36 MIL configurations on 510 WSIs from the UBC-OCEAN dataset using five-fold cross-validation.
- Compared three pathology foundation models (Phikon-v2, CTransPath, UNI), six aggregators (mean/max pooling, ABMIL, CLAM-SB, DSMIL, DTP-TransMIL), and two classification strategies (flat vs. hierarchical).
- Assessed attention map interpretability with pathologist annotations and performed quantitative spatial alignment analysis.
Main Results:
- Feature extractor choice had a greater impact on classification accuracy than aggregator selection.
- Hierarchical cascades consistently improved high-grade serous carcinoma (HGSC) recall across all configurations, achieving 0.949 with UNI max pooling.
- Stronger feature extractors (CTransPath, UNI) produced more spatially structured attention maps compared to Phikon-v2.
Conclusions:
- Feature extractor selection is the most critical component influencing MIL performance in EOC subtype classification.
- Hierarchical classification frameworks offer a significant advantage in improving HGSC detection, particularly when combined with robust feature extractors.
