Deep Learning Model for the Prediction of EBV-associated Gastric Cancer

Abstract

Background: Detection of EBV status in gastric cancer is crucial for clinical decision making, as it identifies patient with different treatment response and prognosis. Despite its apparent importance, EBV status is not readily testable because of limited medical resources. Therefore, widely accessible and cost-effective tools for the testing is required. Here, we investigate the potential of a deep learning-based system for automated EBV prediction directly from hematoxylin and eosin (H&E)-stained whole-slide images (WSIs).

Design: The model was developed using 270 H&E-stained WSIs (24 with EBV positive and 246 with EBV negative), from The Cancer Genome Atlas, containing WSIs. Our deep learning model (EBVNet) consists of two sequential components: a tumor classifier and an EBV classifier (Figure 1). We visualized the features extracted by the tumor classifier (Figure 2A) and the EBV classifier (Figure 2B) using Uniform Manifold Approximation and Projection (UMAP). We internally and externally validated the model on patch-level dataset of TCGA and slide-level dataset of 61 H&E- stained WSIs (6 with EBV positive and 55 with EBV negative). Performance was primary evaluated using the sensitivity, specificity, precision, negative predictive value (NPV) and F1- score.

Results: The EBVNet model achieved a sensitivity of 92.4%, specificity of 99.5%, precision of 92.4%, NPV of 99.5%, and F1-score of 92.4% on the patch-level dataset. On the external dataset (slide-level dataset), our model achieved a sensitivity of 85.7%, specificity of 81.1%, precision of 37.5%, NPV of 97.7%, and F1-score of 52.2%.

Conclusions: Within the current universal EBV testing paradigm, our model might contribute value as to prescreen patients before confirmatory testing, potentially reducing the number of tested patients, thereby resulting in substantial test-related labor and cost savings.