Deep learning model can improve the diagnosis rate of endoscopic chronic atrophic gastritis: a prospective cohort study – BMC Gastroenterology

As a precancerous form of gastric cancer, CAG is a disease of focus that presents difficulties to endoscopists. Early detection and diagnosis of CAG can prevent the formation of gastric cancer to a certain extent, but the difficulty of diagnosis and the rate of missed diagnosis have brought great challenges to endoscopists [11]. China is a country with a large population. Endoscopists often need to perform gastroscopy on 10–20 patients within half a working day and deliver the endoscopic diagnosis report to patients immediately. Therefore, improving the accuracy of endoscopic diagnosis is a challenge that every endoscopists must face. Previous studies have shown that the diagnosis rate of endoscopic CAG varies greatly in different hospitals in the same and different regions, fluctuating from 17.7% to 39.8%, and the sensitivity of endoscopic diagnosis of atrophy is only 42% [2]. Therefore, it is particularly important to improve the diagnosis rate of endoscopic CAG. According to the “Consensus of Chronic Gastritis of China”, the endoscopic manifestations of CAG are red and white mucosa, mainly white mucosa; the folds flatten or even disappear, and some mucosal vessels are exposed, which can be accompanied by mucosal granules or nodules [6]. However, in clinical practice, identifying atrophy is mainly based on the subjective understanding of endoscopists, and it depends on their understanding of the guidelines, previous operating experience, the standard training level of the hospital and other factors. Therefore, there are many uncertainties and great differences. How to perform consistent and accurate early detection and diagnosis of CAG by each endoscopist has always been a difficult problem that clinical guidelines have been attempting, but have been unable, to solve.

The emergence of artificial intelligence, especially DL, has provided a better solution to this problem. DL is currently a research hotspot in the field of machine learning. It has shown excellent performance in image recognition and other fields [3, 12]. Its combined application with digestive endoscopy has become a popular research topic, especially for upper gastrointestinal diseases [13, 14]. Currently, the main research directions include DL auxiliary detection of Barrett's esophagus, esophageal cancer, gastric cancer, Helicobacter pylori infection and anatomical sites, especially for early cancer. In the upper gastrointestinal endoscopy application field, simple reliance on endoscopists for endoscopic diagnosis still has many limitations and difficulties, such as the diagnosis and differential diagnosis of early malignant tumors (early esophageal cancer and gastric cancer, etc.); approximately 10% of malignant lesions might be missing, but computer-aided diagnosis (CAD) can help endoscopists to accurately detect and screen for early cancer. Some scholars have established a CAD system capable of automatic detection of early gastric cancer using a large number of traditional endoscopic pictures and a convolutional neural network in the DL algorithm. It has high speed of lesion recognition and sensitivity of 92%, indicating that the CAD system with this algorithm as the core has strong clinical diagnostic ability [15]. China's independent research and development of cancer of upper gastrointestinal endoscopy AI auxiliary diagnosis system performance are good. A multicenter, randomized, controlled trial indicated that the system of upper gastrointestinal cancer diagnosis has high accuracy, sensitivity and specificity; the sensitivity can be close to the diagnosis of endoscopic experts and has been established as a preliminary application in clinical endoscopic examination [14]. While many scholars have focused on early cancer of the upper gastrointestinal tract, our study focused on early lesions of “early gastric cancer”—“chronic atrophic gastritis” —to “move forward the threshold” and more effectively reduce the occurrence of gastric cancer and monitor the development of gastric cancer. Some studies have achieved high sensitivity recognition of gastric precancerous lesions, such as polyps, ulcers and erosions, by simplifying many parameters of the neural network under the training of a small data set, with sensitivity of up to 88.9% [16]. Research on AI auxiliary detection for the diagnosis of CAG has also gradually developed [17]. A study suggested that the accuracy, sensitivity and specificity of a convolutional neural network for the diagnosis of CAG were 0.942, 0.945 and 0.940, respectively, which were higher than those of endoscopic experts, and the mild, moderate and severe atrophic gastritis detection rates were 93%, 95% and 99%, respectively [18]. However, the data used for training and validation of the model in most of the studies were retrospective, endoscopic static pictures, and the data were artificially preliminarily screened; thus, they lacked prospective research results [19]. Prospective studies have also focused on the recognition of static pictures [20], while the recognition of real-time video monitoring has been less common [17]. Therefore, we conducted a study to apply the DL-based diagnostic model for CAG to real-time video monitoring in gastroscopy. The study suggested that, compared with endoscopist diagnoses, the DL model could effectively improve the diagnosis rate of endoscopic CAG.

Considering that previous studies of CAD-assisted diagnosis of CAG were mostly limited to retrospective static picture recognition studies [17, 18], our study is a good extension of these previous studies. We developed a DL-based diagnostic model for CAG that can be applied for real-time video monitoring in gastroscopy, and we conducted a prospective cohort study, which suggested that this diagnostic model can improve the diagnostic rate of endoscopic CAG compared with endoscopists [35.8% vs. 24.6%, χ2 = 7.962, RR = 1.453(1.117–1.894, P = 0.005]. Endoscopic examination is highly dependent on the clinical experience and status of endoscopists, especially for the endoscopic diagnosis of CAG. Inexperience with endoscopists and fatigue during operations often lead to missed diagnoses [21]. Our DL model can overcome the above shortcomings. For endoscopist, it can serve as an experienced instructor to remind them of the location of atrophy and improve the diagnosis rate and as an assistant who does not get fatigued, reminding the endoscopist in a timely manner about lesions that might be neglected and noting omissions. Therefore, our diagnostic model could be a powerful means to assist endoscopists in the early detection and diagnosis of CAG.

The study also found that the proportions of moderate and severe atrophy patients diagnosed by the DL model were significantly larger than that of the endoscopist. Among the atrophic lesions diagnosed by the DL model, the proportion of severe atrophic lesions was significantly increased, reflecting that the DL model can more accurately and objectively evaluate the severity of atrophy and avoid the misjudgment of endoscopists regarding the patient's condition. Previous studies have shown that the pathological diagnosis of patients with CAG at the second visit is often worse than that at the first visit, which is considered to be related to most patients being screened during the first visit for gastroscopy and few endoscopic precision examinations being performed [22,23,24]. The first visit process of gastroscopy is mainly to find the lesions and perform a preliminary evaluation. The nature and severity of the lesions are mainly determined by pathological diagnosis. Even experienced endoscopists will inevitably misdiagnose lesions. The patients at the second visit were subjected to precision endoscopic examination, which can be evaluated based on the pathological results of the first examination, which is more in-depth and accurate than the first examination. The DL model can effectively avoid the above deficiencies, assisting the endoscopist in labeling all atrophic lesions during the first visit for gastroscopy, accurately assessing the severity of lesions, and performing accurate biopsy.

Among the atrophic lesions diagnosed by the DL model, the proportion of severe intestinal metaplasia was significantly larger than that found by the endoscopist. The range and subtypes of intestinal metaplasia have important value in predicting the risk of gastric cancer, and the severity of intestinal metaplasia determines the assessment of gastric cancer risk staging by Operative Link for Gastric Intestinal Metaplasia Assessment (OLGIM) [11, 25]. The manifestations of intestinal metaplasia under gastroscopy are different, including light yellow nodular type, porcelain white small nodular type, fish-scale type, diffuse type, etc. [2]. These findings are heavily dependent on the operation experience and operation status of the endoscopist, which can lead to missed diagnoses. However, the DL model does not have the above problems, so it can easily and consistently find all types of intestinal metaplastic lesions that the model has learned and evaluate their severity.

In our study, it was found that, compared with endoscopist diagnoses, the proportion of patients diagnosed with “type O” CAG was significantly increased by the DL model, while the proportion of “type C” CAG was not significantly different. Our results were consistent with the conclusions drawn from previous studies. Because the observation of the gastric fundus and gastric body requires turning over the head of the gastroscope, the observation angle and light will be significantly changed compared with frontal observation sites, such as the gastric antrum and gastric angle. Endoscopists must readjust to the angle and light, and any mistake could lead to missed diagnoses [26, 27]. However, the DL model can effectively overcome the above difficulties. The observation angle and light change have almost no influence on its stability, and it can effectively detect atrophic lesions of the gastric fundus and gastric body.

Compared with endoscopist diagnoses, DL model showed that the number of atrophy sites found was significantly increased, the number of biopsies was significantly decreased, and the ratio of the number of atrophy sites found to the number of biopsies was significantly increased. The “Consensus of Chronic Gastritis of China” points out that histopathology is very important for the diagnosis of chronic gastritis, and biopsy should be performed according to the endoscopic conditions and needs. For clinical diagnosis, it is recommended to obtain 2 to 3 pieces of tissue from the gastric antral, gastric angle and gastric body for histopathology, and additional biopsy must be performed of suspicious lesions, the purpose of which is to prevent missed diagnoses through multisite biopsy [2, 28]. However, multisite biopsy is prone to excessive mucosal damage and prolonged operation times, leading to an increased risk of intraoperative and postoperative bleeding and cardiovascular and cerebrovascular complications. The DL model can overcome the above shortcomings, not only improving the number of atrophic lesions detected but also reducing the number of blind-spot biopsies; it not only reduces mucosal injury but also reduces the operation time of gastroscopy. It not only reduced the operation burden and pressure on endoscopists but also reduced the duration of operation-related pain and risks to patients.

We also used the cohort to conduct a nested case–control study. Considering pathological diagnosis as the gold standard, the diagnostic evaluation indices and the evaluation of consistency with pathological diagnosis in the DL group were better than those in the endoscopist group. Therefore, the DL model of CAG has better ability to detect CAG and identify CNAG. Those diagnosed as positive by the DL model have a greater probability of having CAG. Those diagnosed as negative by the DL model have a greater probability of having CNAG. The AUC for the DL model was > 0.9, indicating high diagnostic accuracy. The Kappa value for the DL model was > 0.8, showing good consistency with the pathological diagnosis.

There are some limitations of our study. First, because it was an exploratory study, our exclusion criteria were strict to avoid risks to patients. We excluded patients with endoscopic lesions other than chronic gastritis, such as peptic ulcers and gastrointestinal malignant tumors, as well as patients with chronic medical histories, such as hypertension, diabetes, coronary heart disease and cerebrovascular disease, which resulted in a certain bias in the sample enrolled in the cohort. After obtaining the successful experience of this study, we will include a broader range of patients with chronic gastritis complicated with other endoscopic lesions and with chronic diseases in our future cohorts to further scientifically verify our model. Second, we conducted a nested case–control study with this cohort, considering pathological diagnosis as the gold standard to verify the diagnostic evaluation indices of our model and its consistency with pathological diagnosis. Due to the short study period and insufficiently large cohort, although the sample size met the standards of statistical efficiency, the baseline data of the CAG group and the CNAG group were not completely equal, so the evaluation indices calculated were biased to some extent. In the future, we will continue to expand our cohort, strictly match the baseline data of the CAG group and the CNAG group and recalculate and correct our evaluation indices.. Third, as this is an exploratory study, we conducted the study with a single-center cohort. The enrolled cases were only representative of this region and may have selection bias. In futures stages, we will include additional regions for a multicenter study so as to make our research results more broadly representative.

#Deep #learning #model #improve #diagnosis #rate #endoscopic #chronic #atrophic #gastritis #prospective #cohort #study #BMC #Gastroenterology

Lascia un commento

Il tuo indirizzo email non sarà pubblicato.