Predicting HER2 Expression in Gastric Cancer With a Deep Learning Model
A recent study published in Advanced Science has introduced a deep learning model called HER2Net, which predicts HER2 status in gastric cancer (GC) using hematoxylin-eosin (H&E)–stained pathology slides. HER2 is a key therapeutic target in GC, with anti-HER2 therapies such as trastuzumab offering significant benefit to patients who overexpress the HER2 protein. However, current methods to assess HER2 status, such as immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH), are resource-intensive, subjective, and sometimes inconsistent across samples. HER2Net was developed to offer a more accessible, cost-effective, and consistent approach to HER2 status determination.
The model was trained using 531 H&E whole slide images (WSI) from 520 patients and tested on an internal set of 115 WSI and an external multi-center dataset of 102 WSI. HER2Net achieved an accuracy of 90.4% on the internal test set and 89.2% on the external test set. The architecture consists of 3 components: a tile-level classifier to evaluate small regions of the slide, an integrated classifier to combine tile-level predictions, and a high-expression proportion calculator that determines HER2 status based on the percentage of tumor pixels classified as high-expression. Notably, HER2Net incorporates a pixel-level tumor detector, ensuring that only tumor regions are analyzed, rather than background or nontumor tissue.
In comparison to prior deep learning models applied to GC subtyping, which typically perform binary classification and lack clinical interpretability, HER2Net offers a semi-quantitative output aligned with the ≥10% threshold used in clinical IHC assessments. Importantly, the study identified that misclassifications by HER2Net were sometimes associated with specific histopathological features, such as adenoid or papillary differentiation.
This innovation holds potential as a standalone or adjunctive tool in HER2 diagnostics, especially in settings where access to IHC or FISH is limited. While promising, the authors acknowledge the need for additional validation in larger and more diverse patient cohorts, as well as the opportunity to enhance HER2Net further by integrating newer deep learning architectures.
Reference
Liao Y, Chen X, Hu S, et al. Artificial intelligence for predicting HER2 status of gastric cancer based on whole-slide histopathology images: a retrospective multicenter study. Adv Sci (Weinh). 2025;12(10):e2408451. doi:10.1002/advs.202408451
