Real-World Benchmarking and Validation of Foundation Model Transformers for Endometrial Cancer Subtyping from...

Area of Science:

• Digital pathology
• Computational oncology
• Artificial intelligence in medicine

Background:

• Endometrial cancer (EC) molecular subtyping is crucial for treatment selection.
• Accurate subtyping from whole-slide images (WSIs) using artificial intelligence (AI) can improve diagnostic efficiency.
• Current methods may not fully leverage the potential of advanced AI models for histopathology.

Purpose of the Study:

• To evaluate open-source foundation models with attention-based multiple instance learning (MIL) for classifying EC molecular subtypes from WSIs.
• To assess the performance and generalizability of these models in an independent, real-world cohort.
• To compare foundation models against traditional convolutional neural networks (CNNs).

Main Methods:

• Assembled public (n=815) and external (n=720) EC cohorts with molecular subtyping.
• Benchmarked four CNNs and six open-source foundation encoders using TransMIL and CLAM MIL strategies within the STAMP pipeline.
• Trained models with cross-validation and evaluated on the independent cohort, using macro-area under the receiver operating characteristic curve (AUC) as the primary outcome.

Main Results:

• Foundation models outperformed CNNs in cross-validation (macro-AUC 0.799-0.860 vs 0.715-0.829).
• The best configuration (Virchow2 with CLAM) achieved a cross-validation macro-AUC of 0.860.
• In external validation, foundation models showed higher discrimination (macro-AUC 0.667-0.780) compared to CNNs, with UNI2 and Virchow2 models demonstrating strong performance.

Conclusions:

• Open-source foundation model pipelines with attention-based MIL accurately and generally classify EC molecular subtypes from WSIs.
• These models surpass CNN performance in real-world validation.
• The findings support the use of these AI tools for scalable, cost-effective precision oncology and molecular testing triage.