Hierarchical Vision Transformers for prostate biopsy grading: Towards bridging the generalization gap.

Area of Science:

• Computational pathology
• Medical imaging
• Artificial intelligence

Background:

• Whole-slide images (WSIs) in pathology present computational challenges due to their large size.
• Vision Transformers (ViTs) are powerful but struggle with large-scale image data.
• Hierarchical Transformers offer a solution for processing long sequences or large images.

Purpose of the Study:

• To explore the efficacy of Hierarchical Vision Transformers (HViTs) for prostate cancer grading in WSIs.
• To introduce a novel method for combining attention scores across hierarchical levels.
• To evaluate the model's performance and generalization capabilities.

Main Methods:

• Implementation of a Hierarchical Vision Transformer architecture tailored for WSIs.
• Development of a technique for smart attention score aggregation across transformer hierarchies.
• Validation on the Prostate cANcer graDe Assessment (PANDA) dataset and diverse clinical datasets.

Main Results:

• The best-performing model achieved a quadratic kappa of 0.916 on the PANDA test set, matching state-of-the-art.
• The model demonstrated superior generalization, achieving a quadratic kappa of 0.877 in diverse clinical settings.
• The approach proved robust and practically applicable.

Conclusions:

• Hierarchical Vision Transformers with novel attention mechanisms are highly effective for prostate cancer grading in WSIs.
• The proposed method offers robust performance and excellent generalization, suitable for real-world computational pathology.
• This work paves the way for wider adoption in medical imaging analysis.