Introduction

Chronic inflammation of the gastric mucosa induces a cascade of precancerous conditions (chronic atrophic gastritis, intestinal metaplasia) and lesions (dysplasia) that may result in the development of intestinal-type gastric cancer.1 Infection with Helicobacter pylori and autoimmune gastritis are the most relevant factors initiating these mechanisms. Conventional white-light endoscopy has moderate sensitivity and specificity, as well as a high interobserver variability, and is therefore not sufficient to reliably diagnose gastric atrophy or intestinal metaplasia.2 3 Thus, especially in Western countries, histology-based diagnosis of precancerous conditions using standardised biopsy protocols is favoured. Advanced endoscopic techniques (eg, virtual or conventional chromoendoscopy, magnification endoscopy, confocal laser endomicroscopy) are often hindered by technical availability and costs.

DL has demonstrated potential in medical imaging, including GI endoscopy.4 In this field, DL has been used to diagnose focal pathologies (in particular colorectal polyps and oesophageal adenocarcinoma), and only occasionally for diseases diffusely affecting the GI mucosa (eg, H. pylori-associated gastritis).4–7 Here, for the first time, we present a DL approach that overcomes the limitations of white-light endoscopy in diagnosing atrophic gastritis.

Patients and methods

For a first data set, we identified 200 real-world images from patients with and without histology-proven atrophic gastritis (100 each) from subjects undergoing routine oesophagogastroduodenoscopy between 2008 and 2018 (data set DS1). Endoscopies were performed with various generations of Olympus scopes (GIF-Q160, GIF-Q160Z, GIF-1TQ160, GIF-Q165, GIF-H180, GIF-H190; Olympus Europe, Hamburg, Germany). Images were unaltered white-light images anonymised and exported as Digital Imaging and Communications in Medicine (DICOMs). Non-standardised images (eg, various scope positions, distances, angles and illumination; bile, food and mucus contaminations) were taken from the non-overinflated proximal stomach (gastric corpus and fundus). All images were cropped, resized and normalised to have a set average and SD.

An independent second data set (data set DS2) of 70 images (30 with atrophy; 40 without) was used for independent testing and evaluation by six endoscopists with less than 1500 and more than 1500 esophagogastroduodenoscopy (EGDs). Since the two groups (three each) did not differ (p>0.05), their ratings were combined. Table 1 summarises the patient characteristics. Patients included in the study had no evidence of persisting H. pylori infection. Histopathological assessment of H&E-stained slices was carried out by seven board-certified academic pathologists, with at least two pathologists evaluating each specimen (non-blinded) using the updated Sydney system.8

With traditional machine learning, handcrafted features are fed to a model for classification. With DL these are computed incrementally by the model without expert intervention. Thus, there is no theoretical limit that prevents it from learning any feature representation. Convolutional neural networks (CNNs) are the gold standard for image analysis. CNNs take advantage of the local structural relationships in the image and create progressively more complex abstract representations from layer to layer. However, this requires a large amount of training data.

To overcome this limitation, we used a fine-tuned, pretrained CNN; that is, we used pretrained weights to initialise the network, thus improving the stability and performance of our model (figure 1). The training data were artificially augmented by image rotation, mirroring and scaling. First, we assessed the best architecture (pretrained models on ImageNet).9 We performed 10-fold stratified cross-validation on DS1. For each round, data were split into training, tuning and testing sets (80%/10%/10%). The test set was classified using the best performing hyperparameter combination, as assessed in the tuning set (early stop grid search to select dropout, learning rate, momentum). All images were used for testing once only. In a second stage, DS1 was used for training and tuning (90%/10% split), whereas DS2 was used for testing. Model architecture was chosen from the first-stage results. Hyperparameters were assessed in the tuning set. Online supplementary file 1 provides an indepth description of the methods.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Study workflow. DS1 was used to perform 10-fold stratified cross-validation (A). At each iteration, one fold was used for testing, one was used for tuning, and the remainder were used for training. In a second step, DS1 was used for tuning and training (10/90% split), whereas DS2 was used for hold-out testing (B). The workflow for each training, tuning and testing cycle was performed as presented in C. The testing set (DS2) was only used to derive results and never to train the model or to tune hyperparameters, remaining completely independent. CNN, convolutional neural network; DS1, data set 1; DS2, data set 2.

For all models and expert evaluations, accuracy, balanced accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and F-score were computed. In addition, NPV, PPV and accuracy were computed for prevalence rates between 20% and 50% in steps of 1%. Statistical differences between the expert evaluations and DL were assessed with Wilcoxon signed-rank test. Further, we computed the receiver operating characteristic (ROC) curves and the AUC.

Results

The best performing pretrained DL model for diagnosis of atrophic gastritis, as assessed by cross-validation, was VGG16.10 The algorithm yielded results for all images. Table 2 summarises the results.

Accuracy, balanced accuracy and F-score were significantly lower for the endoscopists when compared with the DL-based approach (p=0.03). There was no significant difference between the endoscopy experts and the model for the remaining performance metrics. Online supplementary figure 1A,B shows the ROC curves.