Diagnostic usefulness of deep learning methods for Helicobacter pylori infection using esophagogastroduodenoscopy images

Abstract

Background and aims: We aimed to assess the diagnostic potential of deep convolutional neural networks (DCNNs) for detecting Helicobacter pylori infection in patients who underwent esophagogastroduodenoscopy and Campylobacter-like organism tests.

Methods: We categorized a total of 13,071 images of various gastric sub-areas and employed five pretrained DCNN architectures: ResNet-101, Xception, Inception-v3, InceptionResnet-v2, and DenseNet-201. Additionally, we created an ensemble model by combining the output probabilities of the best models. We used images of different sub-areas of the stomach for training and evaluated the performance of our models. The diagnostic metrics assessed included area under the curve (AUC), specificity, accuracy, positive predictive value, and negative predictive value.

Results: When training included images from all sub-areas of the stomach, our ensemble model demonstrated the highest AUC (0.867), with specificity at 78.44%, accuracy at 80.28%, positive predictive value at 82.66%, and negative predictive value at 77.37%. Significant differences were observed in AUC between the ensemble model and the individual DCNN models. When training utilized images from each sub-area separately, the AUC values for the antrum, cardia and fundus, lower body greater curvature and lesser curvature, and upper body greater curvature and lesser curvature regions were 0.842, 0.826, 0.718, and 0.858, respectively, when the ensemble model was used.

Conclusions: Our study demonstrates that the DCNN model, designed for automated image analysis, holds promise for the evaluation and diagnosis of Helicobacter pylori infection.

Keywords: Helicobacter pylori; artificial intelligence; deep learning.

Figure 1

Procedure to remove the black region of an image using image processing techniques. (a) Image captured by the endoscopic system. The patient information in (a) is pictorially covered by a red rectangular box. (b) image in which the Canny edge algorithm is applied to the red channel of the color image (a). Morphological operations such as dilation, filling, and complement were used in (c) and (d) to identify corner coordinates in (e) marked with red dots. Using the corner coordinates in (e), the region of interest as shown in (f) was cropped.

Figure 2

Gradient‐weighted class activation map (Grad‐CAM) result of overlaying the heatmap on infected images. The first and third columns show Helicobacter pylori‐infected images. The second and fourth columns show the Grad‐CAM results, which overlie the heatmap on the first and third column images, respectively. Grad‐CAM results were generated using the InceptionResnet‐v2 model. The first‐row images are the true positive images, and the second‐row images are the false negative images. Red areas show highly activated regions, and blue areas depict less activated regions.

Figure 3

Gradient‐weighted class activation map (Grad‐CAM) result of overlaying the heatmap on uninfected images. The first and third columns represent the uninfected images. The second and fourth columns show the Grad‐CAM results, which overlie the heatmap on the first and third column images, respectively. The first‐row images are true negative images, and the second‐row images are false positive images. Grad‐CAM results were generated using the InceptionResnet‐v2 model. Red areas show highly activated regions, and blue areas depict less activated regions.

Figure 4

Receiver operating characteristic curves of deep convolutional neural network and ensemble models. Comparison of receiver operating characteristic (ROC) curves among deep convolutional neural network (DCNN) and ensemble models when all esophagogastroduodenoscopy (EGDS) images were used (upper panel). Comparison of ROC curves of the ensemble model when each sub‐anatomical category EGDS image is used (lower panel). , ensemble (RUC = 00867); , ResNet‐101 (AUC = 0.830); , Xception (AUC = 0.8333); , Inception‐v3 (AUC = 0.834); , InceptionResnet‐v2 (AUC = 0.847); , DenseNet‐201 (AUC = 0.846); , All (AUC = 0.867); , Angle (AUC = 0.753); , Antrum (AUC = 0.842); , cardian and fundus (AUC = 0.826); , lower body GC and LC (AUC = 0.718); , upper body GC and LC (AUC = 0.858).