A histopathology aware DINO model with attention based representation enhancement

Area of Science:

• Histopathological image analysis
• Computational pathology
• Artificial intelligence in medicine

Background:

• Histopathology is crucial for cancer diagnostics.
• Vision Transformers (ViTs) are increasingly used for image analysis.
• Existing attention mechanisms in ViTs may not fully capture histopathological nuances.

Purpose of the Study:

• To introduce HistoDARE (Histopathology-Aware DINO with Attention-based Representation Enhancement), an enhanced attention module for ViTs.
• To improve the extraction of spatially-aware and semantically distinctive features from histopathological images.
• To enhance the performance of cancer detection and classification models.

Main Methods:

• Integrated a novel three-stage AttentionWrapper (spatial, channel, residual refinement) into the DINOv2 ViT-L/14 architecture.
• Applied HistoDARE to the NCT-CRC-HE-100K dataset.
• Interpreted features using Logistic Regression with 5-fold stratified cross-validation.

Main Results:

• Achieved high performance metrics: 98.03% accuracy, 98.03% precision, 98.02% recall, 98.02% F1-score, and 99.95% specificity.
• Outperformed baseline DINOv2 and other state-of-the-art methods.
• Demonstrated significant improvements in clinically critical classes (NORM, STR) and superior class-level consistency.

Conclusions:

• HistoDARE provides a robust and generalizable framework for clinical histopathology.
• The method offers comparable computational efficiency with enriched feature representations.
• HistoDARE strengthens the potential usability of AI models in real pathology workflows.