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 guiding treatment.
• Accurate subtyping from whole-slide images (WSIs) using artificial intelligence (AI) can improve diagnostic efficiency.
• Evaluating the generalizability of AI models in real-world clinical settings is essential.

Purpose of the Study:

• To assess the accuracy of open-source histopathology foundation models with attention-based multiple instance learning (MIL) for classifying EC molecular subtypes from WSIs.
• To evaluate the performance maintenance of these models in an independent, real-world cohort.

Main Methods:

• A discovery cohort (815 patients) and an independent external cohort (720 patients) were utilized.
• Four foundation encoders and four convolutional neural networks (CNNs) were benchmarked using the STAMP pipeline with TransMIL and CLAM MIL strategies.
• Models were trained via cross-validation and tested on the external cohort, with macro-area under the receiver operating characteristic curve (AUC) as the primary metric.

Main Results:

• Foundation models significantly outperformed CNNs in cross-validation (macro-AUC 0.799-0.860 vs 0.715-0.829).
• The best performing foundation model (Virchow2 with CLAM) achieved a macro-AUC of 0.860 in cross-validation.
• In external validation, foundation models demonstrated better performance (macro-AUC 0.667-0.780) compared to CNNs, with UNI2 and CLAM achieving the highest external macro-AUC (0.780).

Conclusions:

• Open-source foundation models coupled with MIL can accurately and generally classify EC molecular subtypes directly from WSIs.
• These AI models surpass traditional CNNs in real-world validation, offering potential for scalable and cost-effective precision oncology tools.
• The findings support the use of these models for guiding treatment decisions and triaging molecular testing in EC.