Multiresolution Self-Supervised Feature Integration via Attention Multiple Instance Learning for Histopathology...

Area of Science:

• Digital pathology
• Computational oncology
• Artificial intelligence in medicine

Background:

• Digital histopathology image analysis is crucial for cancer diagnosis and biomarker discovery.
• Current uniresolution models in deep learning for histopathology have limitations in capturing comprehensive features.
• Patch-based training methods often lose critical information about intratumoral heterogeneity.

Purpose of the Study:

• To develop a multiresolution attention-based multiple instance learning framework for digital histopathology.
• To integrate cellular and contextual features from whole tissue slides for patient outcome prediction.
• To evaluate different methods for combining multiresolution features and compare against uniresolution models.

Main Methods:

• Proposed a multiresolution attention-based multiple instance learning framework.
• Investigated mathematical operations (addition, mean, multiplication, concatenation) for integrating multiresolution features.
• Compared the performance of multiresolution models against uniresolution baseline models for breast-cancer grading.

Main Results:

• All proposed multiresolution models outperformed uniresolution baseline models in breast-cancer grading.
• The multiplication-based multiresolution model achieved the highest performance with an AUC of 0.864.
• Uniresolution baseline models had AUCs of 0.669 and 0.713.

Conclusions:

• Multiresolution analysis is superior to uniresolution approaches for capturing comprehensive features in digital histopathology.
• The developed attention-based multiple instance learning framework effectively predicts patient-level outcomes using whole-tissue information.
• The multiplication-based integration strategy shows significant promise for enhancing prognostic biomarker development in digital pathology.