Image 1_Automatic grading of HER2 immunohistochemical whole-slide images based on a multi-scale hierarchical graph neural network.jpg

Within the detected regions, patch-level HER2 grading integrates Delaunay cell image features with self-supervised visual representations, while WSI-level spatial representations model tumor heterogeneity to achieve fully automated HER2 grading across whole-slide images. The framework was trained using data from two scanning platforms and externally validated on 82 independent WSI samples from a third platform.</p>Results<p>In independent external validation, the Dice coefficient for invasive region detection reached 0.89 (on 50 WSI images); the accuracy rate of HER2 grading at the patch level was 96.49% (95% CI, 96.08–96.87%) (7,924 patches), while that at the WSI level reached 98.0% (95% CI, 91.5–99.9%) (82 WSI images).

The quadratic-weighted Cohen’s κ was 0.97 (95% CI, 0.94–1.00), and all misclassifications at the WSI level were confined to the distinction between HER2 1 + and 2 +.</p>Conclusion<p>This framework enables fully automated HER2 IHC scoring and interpretation under multicenter and cross-scanner conditions, providing a computational tool for standardized IHC assessment, while final HER2 status determination for 2 + cases still requires FISH validation.</p>