This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
The Next Layer: Augmenting Foundation Models with Structure-Preserving and Attention-Guided Learning for Local Patches to Global Context Awareness in Computational Pathology
Created by
Haebom
Author
Muhammad Waqas, Rukhmini Bandyopadhyay, Eman Showkatian, Amgad Muneer, Anas Zafar, Frank Rojas Alvarez, Maricel Corredor Marin, Wentao Li, David Jaffray, Cara Haymaker, John Heymach, Natalie I Vokes, Luisa Maren Solis Soto, Jianjun Zhang, Jia Wu
Outline
EAGLE-Net is a structure-preserving, attention-based architecture based on multi-instance learning (MIL) that overcomes the limitations of conventional foundation models by incorporating mechanisms that leverage both the global spatial structure of the tissue and local contextual relationships between diagnostically relevant regions. It captures global tissue structure through multi-scale absolute spatial encoding, focuses attention on the local microenvironment through top-K neighbor recognition loss, and minimizes false positives through background suppression loss. Evaluated on three cancer type classification tasks (10,260 slides) and seven cancer type survival prediction tasks (4,172 slides) using three different histological foundation backbones (REMEDIES, Uni-V1, and Uni2-h), it achieved up to 3% improved classification accuracy and the highest agreement index in six of the seven cancer types. It also produces smooth, biologically consistent attention maps that align with expert annotations and highlight invasive fronts, necrosis, and immune infiltrates.
Takeaways, Limitations
•
Takeaways:
◦
We present a generalizable and interpretable MIL framework that complements the foundational model to improve understanding of the tumor microenvironment.
◦
Improved prediction accuracy and interpretability are achieved through multi-scale spatial encoding and top-K neighbor-aware loss.
◦
It shows excellent performance in various cancer types and tasks (classification and survival prediction).
◦
Contributes to biomarker discovery and prognostic modeling by generating biologically meaningful attention maps.
•
Limitations:
◦
Since these are performance evaluation results for a specific foundation model and dataset, further research is needed to determine generalization performance for other models or datasets.
◦
There is a lack of comparative performance analysis for models other than the three currently used backbone models.
◦
Validation on a wider variety of cancer types and larger datasets is needed.
