Deep learning-based prediction of HER2 status and trastuzumab treatment efficacy of gastric adenocarcinoma based on morphological features

Abstract

Background: First-line treatment for advanced gastric adenocarcinoma (GAC) with human epidermal growth factor receptor 2 (HER2) is trastuzumab combined with chemotherapy. In clinical practice, HER2 positivity is identified through immunohistochemistry (IHC) or fluorescence in situ hybridization (FISH), whereas deep learning (DL) can predict HER2 status based on tumor histopathological features. However, it remains uncertain whether these deep learning-derived features can predict the efficacy of anti-HER2 therapy.

Methods: We analyzed a cohort of 300 consecutive surgical specimens and 101 biopsy specimens, all undergoing HER2 testing, along with 41 biopsy specimens receiving trastuzumab-based therapy for HER2-positive GAC.

Results: We developed a convolutional neural network (CNN) model using surgical specimens that achieved an area under the curve (AUC) value of 0.847 in predicting HER2 amplification, and achieved an AUC of 0.903 in predicting HER2 status specifically in patients with HER2 2 + expression. The model also predicted HER2 status in gastric biopsy specimens, achieving an AUC of 0.723. Furthermore, our classifier was trained using 41 HER2-positive gastric biopsy specimens that had undergone trastuzumab treatment, our model demonstrated an AUC of 0.833 for the (CR + PR) / (SD + PD) subgroup.

Conclusion: This work explores an algorithm that utilizes hematoxylin and eosin (H&E) staining to accurately predict HER2 status and assess the response to trastuzumab in GAC, potentially facilitating clinical decision-making.

Keywords: Deep learning; Efficacy; Gastric adenocarcinoma; HER2; Trastuzumab.

Fig. 1

Proposed annotation-free deep learning framework for HER2 status classification from WSIs: a, Image segmentation divides images into distinct regions; b, Features are extracted from tissue regions of the WSI using image patches; c, Pre-computed feature vectors of image patches are input into the model. An attention network consolidates patch-level information into slide-level representations, which are utilized for the final diagnostic prediction; d, Strongly patched (red) and weakly patched (blue) regions serve as representative samples to guide clustering layers in distinguishing between positive and negative instances of distinct classes

Fig. 2

AUC results for evaluated weakly supervised methods: a, MIL, b, AttMIL, c, TransMIL, d, CLAM

Fig. 3

Average receiver operating characteristic (ROC) curve generated through the application of stratified 10-fold cross-validation for the surgical cohort (a), including the HER2 2 + dataset (b), with the corresponding ROC curves for the biopsy cohort shown (c); d, The box plot was generated by consolidating the test set scores obtained from each fold for the surgical cohort

Fig. 4

HER2 IHC expression using unannotated slides: a, Results for the 0 and 1+/3 + groups; b, Results for the 0, 1+, and 2+/3 + groups

Fig. 5

a, ROC curve trained using 10-fold cross-validation for efficacy performance; b, Box plot quantified by aggregating the test set scores from each fold

Fig. 6

Heat maps indicating the performance for the different classification algorithms