A New Diagnosis of Intestinal Metaplasia in Gastric Mucosa: Convolutional Neural Network Based on Visual Auxiliary Method

Abstract

Abstract Introduction/Objective Intestinal metaplasia (IM) is a common lesion observed in chronic gastric disease and is thought to be associated with the development of gastric cancer. One of the characteristics of IM is the presence of goblet cells. However, differentiating these inconspicuous goblet cells from glandular cavities and gastric mucus cells using traditional algorithms is inefficient. This paper proposes an interpretable auxiliary diagnosis method for gastric mucosal intestinal metaplasia based on a convolutional neural network(CNN). Methods/Case Report The overall flow chart of the algorithm is shown in Figure 1A. The self-constructed gastric mucosa dataset is utilized in this paper for IM. To improve performance, some modifications are applied to the U-Net. The auxiliary diagnosis process incorporates R34W and R34U for segmenting the glands and IM regions. The R34W represents a dual-resolution input network structure (Figure 1B). For the initial stage of the auxiliary diagnosis, the convolutional neural network calculates patches with both large and small fields for gland region segmentation. Then, the IM glands segmentation network is used to classify regions with and without IM lesions. Results (if a Case Study enter NA) The result of the WSI level mask generated after post-processing is shown in Figure 2A. R34W Net offers significant advantages in differentiating open epithelial structures and proper glands. This network effectively differentiates between proper glands and epithelial regions while maintaining precise segmentation, leveraging high-resolution features in a small visual field. The quantitative indexes of gland segmentation networks are shown in Figure 2B. As part of auxiliary diagnosis, the classification of proper glands with intestinal metaplasia achieves the highest accuracy of 0.95 and AUC of 0.975 in the test datasets (Figure 2C). The area proportion of IM regions generated by CNN is shown in Figure 2D. Conclusion This paper provides a visual auxiliary method based on a CNN for the pathological diagnosis of intestinal metaplasia. The goal is to enhance the speed and accuracy of pathologists in performing this diagnosis. Additionally, we propose a CNN with a dual-resolution input structure, which is utilized for segmenting gastric proper glands.