ABMIL Framework Uses Foundation Models for LUAD Growth Pattern Prediction
A recent investigation introduces an attention-based multiple instance learning (ABMIL) framework aimed at predicting the main growth pattern in lung adenocarcinoma (LUAD) using whole slide images, minimizing the reliance on extensive annotations. This method employs pretrained pathology foundation models as patch encoders, which can be either fine-tuned or kept frozen. Results indicate that fine-tuned encoders enhance performance, with Prov-GigaPath achieving the highest kappa score of 0.699 under the ABMIL framework. In comparison to basic patch-aggregation methods, ABMIL offers more reliable predictions. The findings were published on arXiv (2604.21530).
Key facts
- Lung adenocarcinoma grading depends on identifying growth patterns.
- Common deep learning approaches require extensive annotations.
- ABMIL framework predicts predominant LUAD growth pattern at whole slide level.
- Integrates pretrained pathology foundation models as patch encoders.
- Fine-tuned encoders improve performance.
- Prov-GigaPath achieved highest agreement (kappa = 0.699) under ABMIL.
- ABMIL yields more robust predictions than simple patch-aggregation baselines.
- Published on arXiv with ID 2604.21530.
Entities
Institutions
- arXiv