INTRODUCTION

Helicobacter pylori is the causative organism for the development of chronic gastritis, peptic ulcer, gastric mucosa-associated lymphoid tissue lymphoma, and gastric cancer. H. pylori is categorized as a group I carcinogen by the International Agency for Research on Cancer (1–3). Endoscopy is routinely performed for the evaluation of H. pylori-associated diseases, including large numbers of patients screening for gastric cancer in high-risk areas (4). However, evaluation of H. pylori during endoscopy requires gastric biopsies as endoscopic impression alone is inaccurate, associated with high interobserver variability (2,5).

As a disruptive technology, convolutional neural network (CNN) is increasingly used for facilitating the diagnosis of digestive tract disorders (6–10). CNN establishes a classification protocol by automated learning and extraction of features using images from a large data set (11). Previous studies have reported high accuracy of CNN in diagnosing H. pylori infection, using static archived endoscopic images (12–16). In our previous work, we have developed a panoramic system of gastroscopic images (17), combining CNN and image-guided endoscopy technology to accurately identify gastric landmark classification of precancerous gastric diseases, and H. pylori infection (10,16,18). Subsequently, we have designed a multitask anatomy detection convolutional neural network (MT-AD-CNN) to detect upper gastrointestinal anatomy in real-time (19).

In this study, we aimed to examine the diagnostic accuracy of CNN in real-time evaluation of H. pylori infection during white-light endoscopy (WLE). A computer-aided decision support system based on CNN, CADSS-HP, has been developed for the real-time evaluation of H. pylori infection and was prospectively evaluated in internal and external test cohorts.

METHODS Study design

This was a multicenter study including 3 stages: (i) development of CADSS-HP using archived video recording from Sir Run Run Shaw Hospital (SRRSH), Hangzhou, China; (ii) internal test cohort prospectively recruited from SRRSH; and (iii) external test cohort prospectively recruited from Deqing County People's Hospital, Huzhou, China and Shaoxing People's Hospital, Shaoxing, China. A flowchart of study design was shown (see Supplementary Figure 1, https://links.lww.com/CTG/B29). The study was approved by the ethics committees of the 3 hospitals before initiating (No.20200220-32). The institutional review boards exempted informed consent for patients whose endoscopic videos were preserved in the retrospective database. All participants completed written informed consent before enrolling in the prospective trial. This study was registered with the Chinese Clinical Trials Registry before the initiation (ChiCTR2000030724).

Development data cohort

Patients aged 18–80 years who received WLE and gastric biopsies with available archived video recordings (January 2019–September 2020) at SRRSH and whose H. pylori status confirmed by the urea breath test (UBT) within 3 months before enrollment were collected.

Patients who received H. pylori eradication therapy, history of gastric cancer, history of gastric surgery, or had endoscopic findings of retained gastric food content, mass, and obstruction were excluded. Medical records were reviewed to obtain demographic and clinical data used for the development of CADSS-HP.

Internal and external test cohort

A prospective multicenter test was conducted between August 2021 and August 2022. In addition to the identical inclusion and exclusion criteria of the development cohort, patients who used antibiotics within 1 month or proton pump inhibitors within 2 weeks before endoscopy were excluded. All eligible patients were administered a 13C-UBT before the procedure. Endoscopists, who cannot see the monitor of CADSS-HP, independently evaluate H. pylori status during endoscopy according to the Kyoto classification of gastritis. Gastric biopsies were routinely obtained by endoscopists from, but not limited to, the antrum.

Data preprocessing

The endoscopy examinations were performed with high-resolution WLEs (GIF-H290, GIF-HQ290; Olympus, Japan). Video recordings were used for the development of CADSS-HP. Owing to the focal distribution of H. pylori infection in the stomach (20), manual labeling of the video frames by H. pylori status was not possible. Therefore, a weakly supervised approach was used to classify the videos. First, video data set was preprocessed, including interval sampling, edge border clipping, and effective frame filtering, with the sampling frequency set to 8 Hz. Each image was resized to 256 × 256 pixels to meet ResNet requirement (see Supplementary Figure 2, https://links.lww.com/CTG/B30). Then, each video was converted into a bag data containing about 190 images on average, where each image was considered a discrete point. All bag data were randomly assigned (4:1) to the training and validation sets for the development of CADSS-HP.

Development of CADSS-HP

The CADSS-HP (College of Biomedical Engineering & Instrument Science, Zhejiang University, Hangzhou, China) that uses ResNet34, CNN consisting of 34 layers, was developed. A multi-instance learning algorithm was used to dynamically update the weights of the model and the Top-K infection probability images. The weighted average probability of Top-K infection images was used to predict H. pylori status. Defined threshold for patients with H. pylori infection was set to exceed 0.5. The initial learning rate was set at 0.001, warmed up for 3 epochs, and adjusted according to the CosineAnnealingLR (21) strategy. The optimizer was adaptive moment estimation (Adam) (22). The weight decay coefficient was set at 0.0001, the number of training epochs at 20, and the batch size at 8. Cross entropy was applied as a loss function for it compares the model's predictions with the actual label of the data. The parameters of the backbone network were initialized by Image-Net (23) pretrained model. Furthermore, data augmentations including random horizontal flip, vertical flip, and random rotation between [−30o, +30o] were used to prevent overfitting. The model structure is shown (Figure 1). Detail of weakly supervised H. pylori-infected video classification method is presented (see Supplementary Methods, https://links.lww.com/CTG/B32). Based on previously developed real-time detection techniques (19), CADSS-HP provided real-time output of a continuous number between 0 and 1, indicating the probability of H. pylori infection.

Figure 1.:

The structure of CNN model. A multi-instance learning algorithm was used to dynamically update the weights of the model and the Top-K infection probability images: (i) at the beginning of each training epoch, we use the initialization or the network parameters in the current state to perform a forward calculation on all images and record the Helicobacter pylori prediction probability infection of each image; (ii) sort the instance prediction probabilities under each training bag, select the Top-K images, and construct a new image data set; (iii) taking the H. pylori infection label of the video as the fake label for the Top-K images, calculate the H. pylori infection loss, and perform back-propagation to update the parameters; (iv) the above steps are repeated until the model converges. CNN, convolutional neutral network.

Definition and study endpoints

Histopathology testing (modified Giemsa/immunohistochemistry/methylene blue staining) from gastric biopsies and UBT results were used as the gold standard. H. pylori positive was defined by the presence of H. pylori in histopathology testing from gastric biopsies and/or a positive result of UBT. H. pylori negative required both histopathology testing and UBT to be negative. H. pylori eradication was defined as having received active drug eradication treatment, excluding spontaneous eradication cases. Senior endoscopist was defined by performing more than 3,000 examinations with more than 10 years of experience in endoscopic procedures. Junior endoscopist has less than 5 years of experience in endoscopic procedures or has completed fewer than 1,000 examinations. Gastric atrophy was determined by endoscopists according to the Kimura-Takemoto Classification (24). Primary outcomes were accuracy, sensitivity, specificity, and area under curve (AUC) of CADSS-HP for the evaluation of H. pylori infection.

Sample size

According to the previous study (25,26), we estimated that the prevalence of H. pylori in the Chinese population was 56%, and AUC of diagnosing H. pylori infection was 0.726 for endoscopists and 0.850 for CADSS-HP. A total of 170 patients would be needed with a power of 0.90, a type I error rate of 0.05.

Statistics

Continuous variables were expressed as mean with SD. Receiver operating characteristic curves were plotted, and the AUCs were calculated. The diagnostic property of CADSS-HP was determined using a universal probability threshold of 0.5. Accuracy, sensitivity, and specificity with 95% confidence interval (CI) were calculated by using the Clopper-Pearson method and compared by the 2-tailed 2 sample proportions test. MedCalc 15.0 was used for all statistical analysis. A 2-side P value <0.05 was considered statistically significant.

RESULTS Development of CADSS-HP

During January 2019 to September 2020, 4,110 patients received WLE with gastric biopsies with archived recordings. Five hundred ninety-nine (14.6%) patients with 113,908 images that met the inclusion criteria were used for the development of CADSS-HP (Figure 2). The mean age of the patients was 47.3 ± 13.0, 290 (48.4%) were male, and 155 (25.9%) were diagnosed with atrophic gastritis (Table 1). Furthermore, 349 (58.3%) tested positive for H. pylori infection after 305 patients (50.9%) received UBT addition to gastric biopsies. Finally, video recordings were randomly assigned to 4:1 ratio including 479 (80%) to the training set, whereas 120 (20%) to the validation set.

Figure 2.:

Workflow diagram for the development and test of CADSS-HP. CADSS-HP, computer-aided decision support system for H. pylori infection; CNN, convolutional neutral network; DQCPH, Deqing County People's Hospital; GERD, gastroesophageal reflux disease; SRRSH, Sir Run Run Shaw Hospital; SXPH, Shaoxing People's Hospital; WLE, white light endoscopy.

Table 1. - Baseline characteristics Development data set (SRRSH)
N = 599 Prospective test data set Internal External SRRSH
N = 240 DQCPH
N = 98 SXPH
N = 118 Age, mean ± SD, y 47.3 ± 13.0 45.9 ± 13.6 47.6 ± 13.1 49.5 ± 10.0 Male, % 48.4 41.7 48.0 50.8 Indication for gastroscopy  Gastritis 53.4 41.2 27.5 55.1  GERD 7.0 6.7 3.1 5.1  Screening 17.2 32.1 61.2 35.6  Others 22.4 20.0 8.2 4.2 Atrophic gastritis, % 25.9 27.5 37.8 25.4 Helicobacter pylori infection, % 58.3 46.3 40.8 32.2

Gastric atrophy was determined according to the Kimura-Takemoto Classification. H. pylori infection was defined by the presence of H. pylori in histopathology testing from gastric biopsies and/or a positive result of UBT.

DQCPH, Deqing County People's Hospital; GERD, gastroesophageal reflux disease; N, number of patients; SRRSH, Sir Run Run Shaw Hospital; SXPH, Shaoxing People's Hospital; UBT, urea breath test.

Prospective internal and external test cohort

Between August and December 2021, 757 patients scheduled for endoscopy at SRRSH were recruited to participate in the internal test arm. Between July and August 2022, 436 patients scheduled for endoscopy at 2 centers (Deqing County People's Hospital and Shaoxing People's Hospital) were recruited to participate in the external test arm (Figure 2). Finally, 456 (38.2%) patients including 189 (41.4%) with H. pylori infection from 3 centers were enrolled as test samples to receive endoscopic examinations by 18 endoscopists. The mean age of the patients was 47.2 ± 12.7 years, 207 (45.3%) were male, 133 (29.2%) were diagnosed with atrophic gastritis. 74.6% (340/456) of patients were biopsied from the antrum alone, 13.8% (63/456) from 2 sites (antrum and corpus), and 0.9% (4/456) from 3 sites (antrum, corpus, and angularis). H. pylori infection was determined by positive of gastric histopathology and UBT in 146 (77.2%), gastric histopathology alone in 9 (4.8%), and UBT alone in 34 (18%).

Performance of CADSS-HP

After examining different known series of CNNs for evaluation of H. pylori infection, ResNet34 providing the highest AUC was chosen as the representative model (see Supplementary Table 1, https://links.lww.com/CTG/B31). We used a universal threshold of 0.5 in the development and test sets. In the test set (N = 456), CADSS-HP achieved an AUC of 0.95 (95% CI, 0.93–0.97) with sensitivity and specificity of 91.5% (95% CI, 86.4%–94.9%) and 88.8% (95% CI, 84.2%–92.2%), respectively (Figure 3 and Table 2). The performance of CADSS-HP for diagnosing H. pylori infection in the internal (N = 240) and external test set (N = 216) are shown (Table 3 and Table 4, respectively). A representative video of CADSS-HP on real-timely diagnosing H. pylori infection is shown (Video 1).

Figure 3.:

ROC illustrates the diagnostic ability of CADSS-HP on Helicobacter pylori infection and the prediction of endoscopists in the test set. CADSS-HP provided real-time output of probability of H. pylori infection at each site (antrum, corpus, angularis, and fundus). Weighted mean probability of H. pylori infection in each patient was obtained. The prediction of endoscopists was represented by a green point (N = 456). AUC, area under curve; CADSS-HP, computer-aided decision support system for H. pylori infection; ROC, receiver operating curve.

Table 2. - CADSS-HP vs endoscopists (in the test set) CADSS-HP
N = 456 Endoscopists Total
N = 456 Senior
N = 255 Junior
N = 201 AUC (95% CI) 0.95 (0.93–0.97) — — — Accuracy (95% CI), % 89.9 (86.7–92.5) 83.8 (80.0–87.0)** 85.5 (80.4–89.5) 81.6 (75.4–86.6)** Sensitivity (95% CI), % 91.5 (86.4–94.9) 78.3 (71.6–83.8)** 83.0 (74.5–89.2)* 71.4 (59.8–80.9)** Specificity (95% CI), % 88.8 (84.2–92.2) 87.6 (82.9–91.2) 87.4 (80.6–92.2) 87.9 (80.5–92.8)

AUC, area under curve; CADSS-HP, computer-aided decision support system for H. pylori infection; CI, confidence interval; N, number of patients.

*P < 0.05, **P < 0.01 (compared with CADSS-HP using a 2-tailed 2 sample proportions test).

Table 3. - CADSS-HP vs endoscopists (in the internal test set) CADSS-HP Endoscopists Total
N = 240 Antrum
N = 221 Corpus
N = 221 Angularis
N = 206 Fundus
N = 93 Total
N = 240 Senior
N = 151 Junior
N = 89 AUC (95% CI) 0.94 (0.90–0.97) 0.93 (0.89–0.96) 0.88 (0.82–0.92) 0.83 (0.77–0.88) 0.82 (0.72–0.89) — — — Accuracy (95% CI), % 86.7 (81.7–90.7) 84.2 (78.7–88.7) 78.3 (72.3–83.5) 77.7 (71.4–83.2) 69.9 (59.5–79.0) 76.3 (70.4–81.5)** 78.8 (71.4–85.0)* 71.9 (61.4–80.9)** Sensitivity (95% CI), % 89.2 (82.1–93.7) 84.9 (75.8–90.8) 82.1 (73.4–88.9) 78.4 (69.2–86.0) 91.3 (79.2–97.6) 72.1 (62.8–80.2)** 77.8 (66.4–86.7)* 61.5 (44.6–76.6)** Specificity (95% CI), % 84.5 (77.1–90.3) 83.5 (75.4–89.8) 74.8 (65.8–82.4) 76.9 (67.6–84.6) 48.9 (34.1–63.9) 79.8 (71.9–86.4) 79.7 (69.2–88.0) 80.0 (66.3–90.0)

AUC, area under curve; CADSS-HP, computer-aided decision support system for H. pylori infection; CI, confidence interval; N, number of patients.

*P < 0.05, **P < 0.01 (compared with CADSS-HP using a 2-tailed 2 sample proportions test).

Table 4. - CADSS-HP vs endoscopists (in the external test set) CADSS-HP Endoscopists Total
N = 216 Antrum
N = 216 Angularis
N = 213 Corpus
N = 216 Fundus
N = 143 Total
N = 216 Senior
N = 104 Junior
N = 112 AUC (95% CI) 0.94 (0.90–0.97) 0.91 (0.86–0.94) 0.92 (0.87–0.95) 0.93 (0.88–0.96) 0.92 (0.86–0.96) — — — Accuracy (95% CI), % 93.5 (89.1–96.3) 85.6 (80.1–89.9) 78.9 (72.7–84.0) 78.2 (72.0–83.4) 84.6 (77.4–89.9) 92.1 (87.5–95.2) 95.2 (88.6–98.2) 89.3 (81.7–94.1) Sensitivity (95% CI), % 94.9 (86.9–98.3) 84.6 (74.3–91.5) 90.8 (81.4–95.9) 96.2 (88.4–99.0) 70.0 (55.2–81.7) 87.2 (77.2–93.4) 92.5 (78.5–98.0) 81.6 (66.6–90.8) Specificity (95% CI), % 92.8 (86.7–96.3) 86.2 (79.1–91.3) 72.3 (63.9–79.4) 68.1 (59.6–75.6) 92.5 (84.6–96.7) 94.9 (89.4–97.8) 96.9 (88.2–99.5) 93.2 (84.3–97.5)

AUC, area under curve; CADSS-HP, computer-aided decision support system for H. pylori infection; CI, confidence interval; N, number of patients.

,,]}Performance of endoscopists

Eighteen endoscopists participated in the test, including 9 senior and 9 junior endoscopists. Endoscopists achieved an accuracy of 83.8% (95% CI, 80.0%–87.0%), sensitivity of 78.3% (95% CI, 71.6%–83.8%), and specificity of 87.6% (95% CI, 82.9%–91.2%), respectively. Senior endoscopists evaluated 255 patients with accuracy, sensitivity, and specificity were 85.5% (95% CI, 80.4%–89.5%), 83.0% (95% CI, 74.5%–89.2%), and 87.4% (95% CI, 80.6%–92.2%), respectively. Junior endoscopists evaluated 201 patients with accuracy, sensitivity, and specificity were 81.6% (95% CI, 75.4%–86.6%), 71.4% (95% CI, 59.8%–80.9%), and 87.9% (95% CI, 80.5%–92.8%) (Table 2). The performance of endoscopists in diagnosing H. pylori infection in the internal and external test set are shown (Tables 3 and 4, respectively). Diagnostic performance of an individual endoscopist is shown (see Supplementary Table 2, https://links.lww.com/CTG/B31).

CADSS-HP vs endoscopists

CADSS-HP demonstrated higher accuracy (89.9% vs 83.8%, mean difference = 6.1%, 95% CI 1.6%–10.7%) and sensitivity (91.5% vs 78.3%; mean difference = 13.2%, 95% CI 5.7%–20.7%) compared with endoscopists for evaluation of H. pylori without difference in specificity. Furthermore, CADSS-HP demonstrated higher sensitivity (91.5% vs 83.0%; mean difference = 8.5%, 95% CI 0.4%–17.7%) compared with senior endoscopists for evaluation of H. pylori (Table 2 and see Supplementary Table 3, https://links.lww.com/CTG/B31). In the internal test cohort, CADSS-HP demonstrated higher accuracy (86.7% vs 78.8%, mean difference = 7.9%, 95% CI 0.3%–16.0%) and sensitivity (89.2% vs 77.8%; mean difference = 11.4%, 95% CI 0.65%–23.2%) compared with endoscopic diagnosis by senior endoscopists. In the external test cohort, no differences in diagnostic property were observed between CADSS-HP and endoscopists.

CADSS-HP vs URT/histopathology

The AUC of CADSS-HP for diagnosing H. pylori was inferior to URT (0.90 vs 0.98; mean difference = 0.08, 95% CI 0.04–0.11), and comparable with histopathology (0.90 vs 0.91; mean difference = 0.01, 95% CI −0.03 to 0.04). Sensitivity of CADSS-HP in diagnosing H. pylori was comparable with URT (91.5% vs 95.2%; mean difference = 3.7%, 95% CI −1.8% to 9.4%), better than histopathology (91.5% vs 82.0%; mean difference = 9.5%, 95% CI 2.3%–16.8%), and lower in specificity than URT or histopathology (88.8% vs 100% 99.6%, P < 0.00) (Table 5).

Table 5. - CADSS-HP vs URT/histopathology AUC (95% CI), % P Accuracy (95% CI), % P Sensitivity (95% CI), % P Specificity (95% CI), % P CADSS-HP 0.90 (0.81–0.93) Reference 89.9 (86.7–92.5) Reference 91.5 (86.4–94.9) Reference 88.8 (84.2–92.2) Reference URT 0.980.96–0.99) <0.000 98.0 (96.2–99.0) <0.000 95.2 (90.9–97.7) 0.147 100 (98.2–100) <0.000 Histopathology 0.91(0.88–0.93) 0.733 92.3 (89.4–94.5) 0.200 82.0 (75.6–87.1) 0.006 99.6 (97.6–100) <0.000

AUC, area under curve; CADSS-HP, computer-aided decision support system for H. pylori infection; CI, confidence interval; URT, urea breath test.

DISCUSSION

In this study, we used CNN to develop CADSS-HP, which demonstrated high accuracy in diagnosing H. pylori infection in a prospectively test cohort. To the best of our knowledge, this is the first study evaluating artificial intelligence-guided real-time diagnosis of H. pylori infection during WLE.

H. pylori infection is pathogenic and carcinogenic. Patients with gastric precancerous conditions are recommended to receive testing and treatment, as well as regular endoscopic surveillance (2). Endoscopy is the primary test of evaluating H. pylori-associated disease in areas with high incidence of gastric cancer. Due to the poor sensitivity of WLE to identify H. pylori infection, gastric biopsies are routinely obtained to evaluate for the presence of H. pylori.

In recent years, CNN has made remarkable progress in the field of endoscopy, including gastrointestinal disorders (7,8,27–31). Computer-aided diagnosis studies in upper endoscopy were relatively well established (8,29–31). With derived image data, CNN can facilitate endoscopists with diagnostic and therapeutic interventions by recognizing image features. Given well-described endoscopic features of H. pylori infection (32), the application of CNN was proved promising for evaluation of H. pylori infection. However, previous studies have been limited by the evaluation of archived static images rather than the real-time assessment of H. pylori infection during live endoscopy.

In our study, the sensitivity, specificity, accuracy, and AUC of CADSS-HP based on real-time evaluation of H. pylori infection were 91.5% (95% CI, 86.4%–94.4%), 88.8% (84.2%–92.2%), 89.9% (95% CI, 86.7%–92.5%), and 0.95 (95% CI, 0.93–0.97), respectively. High diagnostic performance remained consistent across internal and external test cohorts from spanning multiple centers, outperforming endoscopist diagnosis. Reasons for the differential performance of endoscopists in the internal and external test cohorts may be related to the length of gastric examination time or experience in using the Kyoto classification. The diagnostic property of CADSS-HP was within the range reported in previous studies (13–15). For example, Nakashima et al (14) developed a CNN model using 162 gastric lesser curvature images to diagnose H. pylori infection demonstrated an AUC of 0.96. Itoh et al (15) performed a CNN-based system using 149 gastric lesser curvature images to diagnose H. pylori gastritis demonstrated sensitivity and specificity of 86.7%. In comparison, our study evaluated in larger sample of patients (113,980 images from 599 patients) and used multiple anatomical locations to comprehensively assess H. pylori infection status. Above studies (13–15) used serology as the reference standard for H. pylori infection, which is not first recommended in clinical practice (2). By contrast, both gastric histopathology and UBT were required for all patients enrolled in the prospective arm of the study to establish an accurate gold standard.

Our findings have clinical implications. UBT and stool antigen test are regarded as the preferred method in the test-and-treat strategy of H. pylori. However, the accuracy of these tests may be altered by use of acid suppressive therapy, antibiotics, and bismuth agents (2,33,34) and does not provide endoscopic lesions. Evaluation of H. pylori infection during endoscopy requires biopsies for histopathology or rapid urease test. High sensitivity (91.5%) of CADSS-HP for real-time diagnosis of H. pylori infection is comparable with biopsy-based testing with reported sensitivity of 88%–92% (35) and URT with a sensitivity of 95.2%, potentially obviating the need for routine gastric biopsies in patients. High sensitivity of 90% in our study also satisfied the acceptable performance thresholds proposed for clinical adoption of new technologies in imaging (36,37). More importantly, CADSS-HP with high sensitivity contributes to guide the endoscopic diagnose of H. pylori infection-related diseases, especially early gastric cancer. Given the high prevalence of H. pylori worldwide (38), our findings might be generalized to a large population undergoing endoscopy, especially those with dyspepsia or screening for early gastric cancer. Furthermore, the advantage of immediate turn-around time for H. pylori status may minimize delayed or missed H. pylori therapy in patients where gastric biopsy was not performed. For example, US studies demonstrated that 19%–52% of patients hospitalized with bleeding peptic ulcers failed to receive H. pylori testing (39,40). Finally, eliminating gastric biopsies to accurately rule out H. pylori infection in patients with low clinical suspicion may substantially affect utilization of resources. For example, a US study estimated charge of $43,073 per diagnosis of H. pylori infection in low (7%) prevalence population for endoscopic evaluation of dyspepsia (41).

We proposed a weakly supervised approach, combined with the CNN technique, achieving real-time diagnosis of H. pylori infection. CADSS-HP was developed using video recordings and prospective validated in different centers, which is more representative of the real-world scenarios and ready for application in clinical practice. In addition, a computer suitable for running CNN models needs to have high-performance hardware devices such as CPUs, GPUs, memory, and storage, whose price generally ranges from several thousand to tens of thousands of yuan. CADSS-HP is affordable for major medical centers, even rural or small ones.

There are limitations in this study. First, CADSS-HP was validated in only 3 centers introducing bias. Center differences likely biased toward lower accuracy. The accuracy of CADSS-HP will also require further validation in other centers including western countries with low-prevalence areas of H. pylori infection. Second, we excluded patients that underwent H. pylori therapy. The incidence of failed H. pylori eradication therapy is becoming common with the emergence of antibiotic resistant strains and highlighting the importance of confirming eradication. Considering the effects of H. pylori therapy on gastric inflammation, CNN trained using data from patients who received prior therapy will be important in clinical practice. Third, compared with other gastric sites, CADSS-HP was not able to reliably recognize valid images of fundus and will require improvement. The reason may partly be explained by limited quality of the endoscopists' manipulation of the fundus during endoscopy, which affects the extraction of image information by the system. Finally, current system could only be used for categorical diagnosis of H. pylori infection. Targeted biopsies of disease lesions associated with H. pylori infections are essential, which involve more complex visual tasks than binary classification models. In the future, CNN is expected to accurately localize H. pylori infection by combining biopsy sites and endoscopic features for training, thus improving the “targeting” of lesion biopsies.

In conclusion, the real-time diagnosis system based on CNN demonstrated high sensitivity for evaluating H. pylori infection during WLE, outperforming endoscopic diagnosis by endoscopists and comparable with URT. The CNN-based diagnosis system may potentially replace gastric biopsies in patients undergoing endoscopy for the evaluation of H. pylori-associated diseases.

CONFLICTS OF INTEREST

Guarantor of the article: Weiling Hu, MD, PhD.

Specific author contributions: Y.Q.S., W.L.H., and J.Q.L.: conceived and designed the study. Y.Q.S., X.W.Z., A.L.C., J.P.W., X.J.W., Z.Z., and A.H.M.: collected the data and performed statistical analysis. X.S.Z. and X.Y.Z.: designed algorithm and trained the model. W.F.Z., Y.C.S., L.Y., and N.L.: contributed to data acquisition and data interpretation. Y.Q.S.: drafted the manuscript. Q.D. and W.L.H.: revised the manuscript. J.J.K.: contributed to the critical and final revision of the manuscript. Y.Q.S., X.S.Z., W.L.H., and J.Q.L.: were the guarantors of this work and take responsibility for the integrity of the data and the accuracy of the data analysis. All authors have access to the study data and review and approve the final manuscript.

Financial support: This work was supported by the Medical and Health Science and Technology Project of Zhejiang Province (grant no. 2020RC064), the Key R&D Program of Zhejiang Province (grant no. 2021C03111), and the National Natural Science Foundation of China (grant no. 81827804 and no. 31771072).

Potential competing interests: None to report.

Ethics committee: The study was approved by the Ethics Committees of the 3 hospitals before initiating (No. 20200220-32).

Clinical trial: This study was registered with the Chinese Clinical Trials Registry prior to the initiation (ChiCTR2000030724).

Study Highlights

WHAT IS KNOWN ✓ Evaluation of H. pylori during endoscopy requires gastric biopsies as endoscopic impression is inaccurate. ✓ CNN has been applied for the image-based evaluation of H. pylori infection.

WHAT IS NEW HERE ✓ A system utilizing CNN can diagnose H. pylori infection in real time under white light endoscopy. ✓ CNN demonstrated high sensitivity in diagnosing H. pylori infection in the real-time test, outperforming endoscopic diagnosis by endoscopists and comparable to URT. REFERENCES 1. Parsonnet J, Friedman GD, Vandersteen DP, et al. Helicobacter pylori infection and the risk of gastric carcinoma. N Engl J Med 1991;325(16):1127–31. 2. Malfertheiner P, Megraud F, O'Morain CA, et al. Management of Helicobacter pylori infection-the Maastricht V/Florence consensus report. Gut 2017;66(1):6–30. 3. Ogura K, Hirata Y, Yanai A, et al. The effect of Helicobacter pylori eradication on reducing the incidence of gastric cancer. J Clin Gastroenterol 2008;42(3):279–83. 4. Laine L, Cohen H, Sloane R, et al. Interobserver agreement and predictive value of endoscopic findings for H. pylori and gastritis in normal volunteers. Gastrointest Endosc 1995;42(5):420–3. 5. Redeen S, Petersson F, Jonsson KA, et al. Relationship of gastroscopic features to histological findings in gastritis and Helicobacter pylori infection in a general population sample. Endoscopy 2003;35(11):946–50. 6. Cho BJ, Bang CS, Park SW, et al. Automated classification of gastric neoplasms in endoscopic images using a convolutional neural network. Endoscopy 2019;51(12):1121–9. 7. Urban G, Tripathi P, Alkayali T, et al. Deep learning localizes and identifies polyps in real time with 96% accuracy in screening colonoscopy. Gastroenterology 2018;155(4):1069–78.e8.