Combination of artificial intelligence‐based endoscopy and miR148a methylation for gastric indefinite dysplasia diagnosis

Abstract Background and Aim Gastrointestinal endoscopy and biopsy‐based pathological findings are needed to diagnose early gastric cancer. However, the information of biopsy specimen is limited because of the topical procedure; therefore, pathology doctors sometimes diagnose as gastric indefinite for dysplasia (GIN). Methods We compared the accuracy of physician‐performed endoscopy (trainee, n = 3; specialists, n = 3), artificial intelligence (AI)‐based endoscopy, and/or molecular markers (DNA methylation: BARHL2, MINT31, TET1, miR‐148a, miR‐124a‐3, NKX6‐1; mutations: TP53; and microsatellite instability) in diagnosing GIN lesions. We enrolled 24,388 patients who underwent endoscopy, and 71 patients were diagnosed with GIN lesions. Thirty‐two cases of endoscopic submucosal dissection (ESD) in 71 GIN lesions and 32 endoscopically resected tissues were assessed by endoscopists, AI, and molecular markers to identify benign or malignant lesions. Results The board‐certified endoscopic physicians group showed the highest accuracy in the receiver operative characteristic curve (area under the curve [AUC]: 0.931), followed by a combination of AI and miR148a DNA methylation (AUC: 0.825), and finally trainee endoscopists (AUC: 0.588). Conclusion AI with miR148s DNA methylation‐based diagnosis is a potential modality for diagnosing GIN.

Endoscopic screening in Asian countries has reduced gastric cancer mortality. 2 Early diagnosis increases the five-year survival rate to >90%. However, early gastric cancer may be difficult to endoscopically diagnose (4.6-25.8% false-negative rates). [3][4][5][6] Candidate molecular markers (ie, methylation, mutation, and microsatellite instability [MSI]) have been reported as accurate markers to detect early gastric cancer. Artificial intelligence (AI)-based endoscopy has vast medical applications.
We aimed to compare and evaluate the diagnostic sensitivity and specificity of physician-performed endoscopy, AI-based endoscopy, and/or molecular markers in detecting gastric indefinite dysplasia (GIN).

| Candidate molecular marker analysis
Bisulfite polymerase chain reaction (PCR) and quantitative promoter DNA methylation analysis of pyrosequencing of candidate genes (BARHL2, MINT31, TET1, miR-148a, miR-124a-3, NKX6-1) were performed using an EpiTect Bisulfite kit (QIAGEN) and Pyromark Advanced Q24 system (Qiagen). Pyrosequencing quantitatively measures the methylation status of several CpG sites in each gene promoter. These adjacent sites usually show highly concordant methylation patterns. Therefore, the mean percentage of methylation at the detected sites was used as a representative value for the gene promoter. All primers and protocols were based on previous reports. 9-14 TP53 mutation analysis was performed and decided using immunohistochemistry (Anti-TP53 [DO-7] antibody). MSI analysis also was performed as previously described. 15

| Constructing a convolution neural network (CNN) algorithm
To construct an AI-based diagnostic system, we used a deep neural network architecture called the single-shot multi-box detector (SSD) (http s://arxi v.org/abs/1512.0232 5), without altering its algorithm.
SSD is a deep CNN that consists of ≥16 layers. The Caffe deep learning framework was then used to train, validate, and test the CNN. All CNN layers were fine-tuned using stochastic gradient descent with a global learning rate of 0.0001. Each image was resized to 300 pixels ´ 300 pixels. The bounding box was also resized accordingly to optimize CNN analysis. These values were set up by trial and error to ensure that all data were compatible with SSD (AI Medical Service Inc.).

| Outcome measures of AI-based detection and diagnosis
A total of 2961 images from 32 cases were collected. There were some images that were unsuitable for consideration (GIN lesion was not in the image, out of focus images, blurry images, halation affect).
Thus, two expert (board-certified) endoscopist carefully selected 248 images finally that showed the GIN lesion without any issues.
ESD in all cases was performed by two expert (board-certified) endoscopist. When the CNN detected a gastric mucosal abnormality in the lesion for all images, the CNN encodes a disease name (nontumorous lesion or tumorous lesion [cancer and/or adenoma]) and its position. The detected lesion was identified by a yellow rectangular frame on the endoscopic images, and the degree of reliability was calculated according to the measured result of the CNN.

| Statistical analysis
All statistical analyses were performed using the SPSS for Windows All reported p-values were two-sided, and statistical significance was set at p < 0.05. We computed the median DNA methylation value and range for each sample, and we defined the receiver operating characteristic (ROC) curve using SPSS. The z-score analysis was used to normalize the methylation levels of several genes, microsatellite instability (MSI), TP53 gene mutation status, AI diagnosis, and endoscopist's diagnosis in each sample. The z-score of the methylation for each gene was calculated as follows: z-score = (methylation level of each sample -mean value of methylation level)/standard deviation of methylation level. In this analysis, a z-score >0 indicates that the methylation level is greater than the mean value for the population.

| Clinical characteristics and endoscopic images of all cases, lesions
All endoscopic images of abnormal (indeterminate) lesions from the 32 cases of ESD histopathologically confirmed GIN were analyzed for AI diagnosis. Examples of the endoscopic images are shown in Figure 1A. The clinical characteristics of all cases are shown in Figure 1B and Table 1.

| Clinical diagnostic ability of AI, molecular markers, and endoscopist
We first calculated the individual diagnostic ability of AI, molecular markers, and two endoscopists using the area under the curve (AUC) of the ROC curve in Figure 2.
The best AUC was for the biomarker miR148a DNA methylation Mut shown in Figure 2B. Interestingly, the diagnostic power of the AI + miR148a combination was located between the diagnostic power of board-certified endoscopists (AUC: 0.931) and trainees (AUC: 0.588) ( Figure 2C).

| DISCUSS ION
Gastric cancer has the sixth highest incidence and the third most common cause of death from cancer worldwide. However, early gastric cancer may be difficult to endoscopically diagnose. Candidate molecular markers (ie, methylation, mutation, and MSI) have been reported as accurate markers to detect early gastric cancer.
Image recognition using AI has dramatically improved due to innovative technologies such as machine learning and deep learning. These techniques are now being applied to gastrointestinal endoscopy worldwide. AI has high diagnostic accuracy for esophageal, gastric, and colorectal cancers. [16][17][18] However, AI has been mostly used to identify irregular or malignant lesions. Qualitative investigations for a comprehensive diagnosis to facilitate appropriate therapy remain limited. Despite direct visualization of the lesion during endoscopy, findings can be difficult to classify as benign irregular lesions (ulcers, infections, or other factors) or malignant even in using additional narrow band imaging, red dichromatic imaging, extended depth of field, and magnified view functions ( Figure 1C).
Post-endoscopy, biopsy remains essential for clinicians for a definitive diagnosis and to formulate a treatment plan. 19 Despite the availability of both endoscopic and histologic diagnosis, differentiating between benign and malignant lesions is still challenging; some lesions are classified indeterminately as GIN. 20 One reason is the difficulty of evaluating the entire lesion pathologically using only a fragment of tissue. In some cases, patient consent is obtained to perform minimally invasive ESD for both therapeutic and diagnostic purposes. [21][22][23] This makes it possible to perform a histological assessment of the entire lesion to determine whether it is benign or malignant. Conventionally, it is ideal to assess ESD.
Mechanisms of malignant transformation due to genetic abnormalities include driver gene mutations and the accumulation of passenger gene abnormalities due to epigenetic alterations including DNA and microRNA methylation. [24][25][26] In gastric cancer, the accumulation of abnormalities due to epigenetic gene alterations from Helicobacter pylori infection is important in tumorigenesis. 27 Multiple studies have investigated the clinical applications of these candidate molecular markers for diagnosis; while the "risk and predictive diagnoses" for malignant gastric transformation has been achieved, its practical applications to diagnose the presence, site, and extent of lesions have not been sufficiently realized.
AI diagnosis, which involves assessments based on various data on the surface of the lesion, is gaining popularity because it enables the non-invasive diagnosis of the presence, site, and extent of lesions ( Figure 1C). The diagnostic capacity of AI for gastric cancer has been reported superior to that of endoscopists in training, and at par with specialists (board-certified endoscopic physicians). We used AI diagnosis alone or in combination with molecular markers (methylation, mutation, MSI) and endoscopic diagnosis in 32 patients who underwent ESD with a preoperative pathological diagnosis of GIN to perform retrospective single factor and multifactor assessments with ROC. We observed that the accuracy for GIN diagnosis from the combination of miR-148a and AI was extremely high. The AUC results were second only to certified endoscopists for diagnosing GIN lesions. This high accuracy may be due to the combined benefits of AI (surface information) in its ability to diagnose the presence, site, and range of the entire lesion and the accuracy of the molecular marker (cell information). This overcomes the limitation of histologic diagnosis that uses only limited tissue samples ( Figure 1C).
Digital technology is already expanding in the medical field, even in the clinical diagnosis of gastric cancer. Clinicians often use endoscopy for diagnosis not only with white light, but also with a digital magnified function and narrow banding imaging function. Indigo carmine staining is also a helpful tool for diagnosis; however, it does not have enough diagnostic power for gastric cancer, especially for tiny lesions and any artifact lesions (biopsy scar, H. pylori infection, ulceration, inflammation, or drug effect) even when used by an expert endoscopist. AI diagnosis may have the potential to support endoscopists of all skill levels and may also be helpful to shorten the learning curve for trainee endoscopists. Moreover, we speculate that a combination of molecular markers may not only be useful for a more detailed assist diagnosis, but may also be an auxiliary tool for therapeutic strategies.
Our study has several limitations. This study only examined a relatively small sample size using a limited number of genetic markers.
Going forward, more studies involving a comprehensive gene search are needed.