Attentionベースの表現強化を備えたヒストパソロジー認識DINOモデル

Histopathological image analysis plays a critical role in modern medical diagnostics, particularly in the detection and classification of various types of cancer. This study proposes a method called HistoDARE (Histopathology-Aware DINO with Attention-based Representation Enhancement), which offers an innovative approach to the attention module used in the Vision Transformers architecture. Unlike conventional attention mechanisms, HistoDARE introduces a novel three-stage AttentionWrapper module that sequentially applies spatial and channel attention followed by a residual refinement stage, enabling the extraction of spatially-aware and semantically distinctive feature representations. HistoDARE is a method integrated into the DINOv2 model, which uses the ViT-L/14 architecture. The obtained features were interpreted using Logistic Regression, and 5-fold stratified cross-validation was applied on the NCT-CRC-HE-100K dataset. The proposed HistoDARE achieved a mean accuracy of 98.03%, precision of 98.03%, recall of 98.02%, F1-score of 98.02%, and specificity of 99.95%, outperforming the baseline DINOv2 and other state-of-the-art methods. The experiments were conducted on a computer with high computational capacity. Based on the DINOv2 architecture, the proposed HistoDARE maintains comparable computational efficiency and resource usage while generating more contextually enriched and discriminative feature representations. During performance measurements, it demonstrated consistent and stable improvements across all stages in all folds. Notably, significant performance improvements were achieved in clinically critical classes such as NORM and STR. These results demonstrate that HistoDARE not only achieves high overall accuracy but also provides superior class-level consistency, making it a robust and generalisable framework for clinical histopathology applications. The developed method has been shared on our GitHub repository. This ensures transparency in terms of reproducibility and supports its usability by other researchers on different datasets in the future. The core contribution of HistoDARE is a three-stage AttentionWrapper (spatial, channel, residual refinement) integrated into the DINOv2 ViT-L/14 backbone to make patch-level representations histopathology-aware. Despite the small numerical gain over a strong self-supervised baseline, this attention-enabled refinement yields statistically consistent improvements on clinically sensitive classes (NORM, STR) and thus strengthens the model's potential usability in real pathology workflows.