No authorization, no reproduction of this journal article
[Quote this article]
Wang Police, Zhu Yizhi, Wu Hui, He Xinqi, Dong Zehua, Huang Man Ling, Chen Yisi, Liu Mun, Xu Qinghong, Yu Honggang, Wu Qi. Influence of artificial intelligence on the identification capacity of gastric cancer undernectoscopic inner mirror of endoscopic physician [J].Chinese Digestive Endoscopy Magazine, 2021, 38 (10): 783-788.
Wang Jing, Zhu Yijie, Wu Lianlian, He Xinqi, Dong Zehua, Huang Manling, Chen Yisi, Liu Meng, Xu Qinghong, Yu Honggang, Wu Qi. Influence of artificial intelligence on endoscopists’ performance in diagnosing gastric cancer by magnifying narrow banding imaging [J] .CHIN J DIG Endosc, 2021, 38 (10): 783-788.
DOI: 10.3760 / cma.j.cn321463-20210110-00020
Wang Police 1 Zhu Yizhen 2 Wu Lianzhang 2 He Xinqi 2 Dong Zhuhua 2 Huang Man Ling 3 Chen Yisi 4 Liu Meng 5 Xu Qinghong 6 in Hong Kong 2 Wu Qi 1
1 Key Laboratory of Malignant Tumors, Ministry of Transformation and Research, Ministry of Transformation, Department of Western, Beijing University Cancer Hospital and Beijing Cancer Prevention and Treatment, Ministry of Transformation;
2 Wuhan University People’s Hospital Gastroenterology Digestive System Disease Hubei Province Key Laboratory Hubei Province Digestive Disease Miniature Treatment Medical Clinical Research Center;
3 Wuhan Central Hospital Gastroenterology;
4 Wuhan First Hospital Gastroenterology;
5 Wuhan Third Hospital Gastroenterology;
6 Wuhan Eighth Hospital Digestive Endoscopic Center
Wang Police and Zhu Yizhen have equal contributions to this article
Author: Wu Qi
PurposeEvaluating Artificial Intelligence (AI) Auxiliary Gastric Cancer Diagnosis System In real – time dyeing enlarged endoscopic video on endoscopic physicians identify gastric cancer capacity.methodRetrospective Collection March 2017 – January 2020 Wuhan University People’s Hospital and public data concentration early gastric cancer and non-cancer dyed amplifying endoscopic images as training sets and independent test sets, including 4 667 pictures (1 950 Zhang Eastern gastric cancer and 2 717 non-cancer), the test set included 1 539 pictures (483 high-speed gastric cancer and 1 056 non-cancer). Model training in deep learning. Prospective Collection June 9, 2020, November 17, 2020, from Peking University Cancer Hospital and 100 patients of 100 patients of Wuhan University People’s Hospital (including 38 cancer and 62 cases of non-cancer) as a video test set. Incorporating 4 different annular phrases from another 4 hospitals, diagnose video test sets twice (no or AI assist), assessing the impact of AI to judge the ability of gastric cancer in endoscopic physicians.resultWhen there is no AI auxiliary, endoscopic physicians diagnose video test concentration accuracy, sensitivity and specificity of 81.00% ± 4.30%, 71.05% ± 9.67%, and 87.10% ± 10.88%; Under Ai assist, endoscopy Identify the accuracy, sensitivity and specificity of gastric cancer, 86.50% ± 2.06%, 84.87% ± 11.07% and 87.50% ± 4.47%, diagnostic accuracy (P=0.302) and sensitivity (P=0.180) more no AI There is improved when assisting. The accuracy of the gas cancer is 88.00% (88/100) in the video test concentration, and the sensitivity is 97.37% (37/38), the specificity is 82.26% (51/62), and the sensitivity of AI is higher than the internal mirror physician. The average level (P=0.002).in conclusionThe AI-auxiliary diagnostic system is an effective tool to assist in the diagnosis of gastric cancer in dyeing amplifying endoscopic mode, which can improve the diagnostic ability of endoscopic physicians on gastric cancer. It can remind endoscopic physicians to pay attention to high-risk regions to reduce missed diagnosis.
[Key words] artificial intelligence; gastric cancer; narrowband light imaging
Fund Project: National Natural Science Foundation (81672387); Capital Health Development Scientific Research Special (2020-2-2155); Hubei Digestive Disease Micro Immediate Treatment Medical Clinical Research Center (2018BCC337); Hubei Province Major Science and Technology Innovation Project (2018-916 -000-008)
Gastric cancer is the third most common cancer in the world, accounting for the third place for cancer death. The five-year survival rate in progress in the stomach cancer is only 30%, while the five-year survival rate of early gastric cancer is as high as 90%. The gantry diagnosis of gastric cancer at home and abroad pointed out that it is important to reduce early gastric cancer misses an important role in improving the five-year survival rate and improving prognosis. Endoscopy is an effective means of discovering early gastric cancer. In my country, the diagnosis and treatment rate of early gastric cancer is less than 10%, and we have a serious situation in the diagnosis of early gastric cancer. Inventions of Narrow Band Imaging, NBI endoscopic endoscopic, improved gastric cancer detection rate. However, due to the end of the endoscopic physician, the accuracy of the diagnosis of gastric cancer in dyeing amplification mode is also uneven.
Effective learning or accessories may increase the early gastric cancer detection rate of endoscopic physicians. In recent years, artificial intelligence (AI) technology has advanced in the field of digestive endoscopy. In the study of AI-assisted diagnosis in narrowband photographic mode, the accuracy of the model reached 90.91%. However, the past studies are mainly in a still picture centralized verification model. In the previous study, the early gastric cancer white light mirror assist diagnosis system improved the detection rate of early gastric cancer under white light mode. After training and improvement, we have established an endoscopic auxiliary diagnostic system for detecting gastric cancer in dyeing amplification mode. This study is intended to assess the ability of the system in clinical practice to identify gastric cancer.
Materials and Methods
First, the training and the establishment of model
1. Establishment of the training and test sets still pictures: a retrospective collection March 2017 – January 2020 from People’s Hospital of Wuhan University or public data sets 1 811 cases, a total of four 667 stained larger image training set comprising 1 early 950 (1,042 cases) of gastric and 2 717 (769 cases) noncancerous image. Another 1539 images as an independent test set from People’s Hospital of Wuhan University, including 483 early gastric cancer images from 92 patients and one 056 images from 161 cases of non-cancer patients.
All images pathology as the gold standard. Exclusion criteria were: inadequate preparation before the test; under 18 years of age. The image is collected under endoscopic image of narrow-band light imaging mode (using the Japanese company Olympus CV-290 series master camera), all images are cut off black edges, hide patient information.
2. Construction of staining gastric recognition model under magnification mode: real-time diagnosis of gastric cancer classification model is constructed by two deep convolutional neural network, determines the stomach real endoscopic image data in the case of cancerous or non-cancerous. Construction of the model identified by gastric Resnet-50. Use transfer learning to train the model.
Second, the clinical validation
1. Establish Video Test Set: Prospectively collected 2020 June 9 – 17 November 2020 endoscopic imaging data from 94 patients Peking University Cancer Hospital and six patients from the People’s Hospital of Wuhan University. 38 cases where cancer, including 36 cases of early gastric cancer, 2 cases of advanced gastric cancer; non-cancerous 62 cases. 32 cases of early gastric cancer from Peking University Cancer Hospital, four cases of early gastric cancer and two patients with advanced gastric cancer from the People’s Hospital of Wuhan University.
Endoscopic image fragments comprising fragments and staining lesions enlarge white mode, in which white light segment length (10.26 ± 3.32) s, staining when amplified fragment length (58.66 ± 16.50) s. AI model dyeing enlarged fragmentary judge, as an additional white clip information provided to physician involved in testing endoscopic assisted determination. Video clips are reviewed by experts and clip lesions fragments are faded patient information. Video clips inclusion criteria: Age ≥18 years of age; there is a clear focus needs to be painless staining magnifying endoscopy examination to further clarify the characteristics of the lesion. Exclusion criteria: history of previous gastrectomy, gastric remnant; without pathological diagnosis; active bleeding lesions observed impact; white fur covered surface of the lesion, the flushing difficulties, observe the effect; for other reasons observe the effect of diagnosis, as previously biopsies; researchers believe that the subject is not suitable to participate in other cases in this study. The test set is not coincident with the above-mentioned training set and test set still pictures, is an independent test set. The same test cases to pathology results as the gold standard, subject to see the video after the start of the test, you can not watch the video in advance. Image data are using the Japanese company Olympus CV-290 Series host shooting.
2. endoscopists test: There are four different years of endoscopists from four hospitals to participate in this verification, including a high qualification (endoscopic experience of more than 10 years), MD, a middle-aged capital (endoscopic experience 5 to 10 years) physicians and 2 junior (less than 5 years of experience in endoscopy) physician. Endoscopists inclusion criteria: agree to participate; there is more than one year of experience operating independently endoscopy, gastroscopy cumulative amount> 500 cases; there is staining magnifying endoscopy experience, once used independently staining magnifying endoscopy in patients with lesions properties of endoscopic diagnosis.
Physicians involved in a total of 2 test, the first test no AI prompt, 2nd test has prompted AI, 2 tests at 2 week intervals (as a forgotten period). When tested, the participants in the same classroom, within the specified time (3 h) with a unified computer test video, judge each case is cancerous or non-cancerous. In the case of video playback, recording participants believed the extent (50% to 100%) when their judgment and judgment. The standard answer will be announced after the completion of 2 tests.
Third, statistical methods
Using SPSS 26.0 statistical software for data analysis through accuracy, sensitivity, specificity, positive predictive value and negative predictive value to assess the performance of endoscopists and AI; using McNemar’s test and chi-square test to compare the endoscopist in the presence or absence differences in the difference between them and an auxiliary AI and AI; conformability between Kappa test subject analysis, Kappa value of 0 to 0.20 as the low consistency,> 0.20 to 0.40 as the general consistency,> 0.40 to 0.60 for medium consistency,> 0.60 to 0.80 for the high consistency,> 0.80 to 1 for the consistency is very high. P <0.05 was considered statistically significant differences.
Participating in the study signed the informed consent of all endoscopists. This study by the People’s Hospital of Wuhan University (approval number: WDRY2019-K094) and Peking University Cancer Hospital (approval number: 2020YJZ08) Ethics Committee approval.
A, AI in pictures and videos of the performance test set
AI 1 539 identify images in the test set accuracy was 91.62% of gastric cancer (1 410/1 539), the sensitivity and specificity of 91.93% (444/483) and 91.48%, respectively (966/1 056). Video test concentration in 100 cases, the accuracy of the AI ??was 88.00% (88/100), to identify whether the sensitivity of cancer patients was 97.37% (37/38), specificity was 82.26% (51/62).
Two, AI assisted endoscopic diagnosis of gastric cancer physician capacity and efficiency
Endoscopist and (or) the video test set AI expression in Table 1, showing: AI at endoscopist assistance, recognition accuracy of gastric cancer (χ2=1.350, P=0.302), sensitivity (χ2=1.953, P=0.180) have a certain upgrade, significantly higher than the sensitivity of AI AI assisted endoscopist (χ2=9.896, P=0.002). Typical Legend Figure 1.
figure 1Artificial intelligent auxiliary diagnostic system identifies a typical example of gastric cancer in real-time staining enlarged video1A: Artificial intelligence tips did not detect cancer;1b: Artificial intelligence reminder suspected cancer
The performance of each participant is shown in Figure 2. It can be seen that the low-income physician has a significant increase in accuracy (χ2=7.579, p=0.004) and specificity (χ2=5.818, p=0.012); Middle-aged physicians have significantly improved under AI assistance (χ2=4.000, p=0.039).
figure 2Different year endoscopic physicians diagnosed gastric cancer performance during video test concentration2A: Diagnostic accuracy comparison;2b: Diagnostic sensitivity contrast;2C: Diagnostic specificity comparison;2D: Diagnosis time comparison
Third, the consistency of gastric cancer before and after AI auxiliary, different year
High-year physicians have a high degree of expression before and after AI (kappa=0.702), and the presence of the middle-aged physician has also shown high levels (kappa=0.758). Low-year physicians have high comparative comparison and high (kappa=0.620, kappa=0.605). The consistency between 2 low-year physicians appeared as moderate consistency before AI (kappa=0.460), and the AI ??auxiliary was increased (kappa=0.626).
This study built a gastric cancer auxiliary diagnostic system in endoscopic amplifier mode. The system assisted endoscopic physician determines whether there is a full assessment of gastric cancer lesions in video testing. Under the aid of this system, the diagnostic level has improved, and the diagnostic time is less. This reflects the effectiveness of the system to judge the validity of gastric cancer in dyeing amplification mode. The system can remind endoscopic physicians to observe the possibility of cancer-resistant areas, and give recommendation for lunar dedication, which can improve the diagnostic efficiency of endoscopic physicians and reduce missed diagnosis.
In recent years, with the rapid development of the AI ??in the field of digestive endoscopes, many scholars have carried out AI auxiliary diagnosis related to gastric cancer. Among them, there are also studies to study the early gastric cancer identification model under narrowband light imaging mode, and verify the effectiveness of the AI ??system in the picture test. In real time, this study verified the effectiveness of the AI ??auxiliary diagnosis of gastric cancer models, which were possible for the AI ??auxiliary diagnostic system.
The gastric cancer auxiliary diagnostic system in the inner mirror staining amplification mode proposed in this study provides an effective auxiliary tips in the process of judging the gastric cancer. The overall performance of AI is better than 4 internal mirrors, and AI judges that the sensitivity of gastric cancer lesions is significantly higher than the average of endoscopy in dyeing amplification mode. In addition, after 2 weeks of forgetting, 4 participated endoscopic physicians raised the sensitivity of gastric cancer lesions in AI aid, and the accuracy rate increased by the first test, and the consistency between low-year physicians also increased. Explaining that the internal mirror is in the help of the auxiliary diagnostic system, the ability to diagnose gastric cancer has improved, especially for low-year-old doctors, the system is a tool worthy of reference in clinical practice.
The AI ??auxiliary diagnostic system can be used as a convenient and reliable tool. Endoscopic physicians have a longer than the end of the inner mirror physician at the end of the inner mirror. Note The participant’s efficiency is higher under AI auxiliary diagnosis. The high sensitivity of AI in the test concentration proves that the system will hardly miss gastric cancer. It has a good performance in the static and dynamic independent test concentration of the two hospitals, indicating that the auxiliary diagnostic system has good stability and generalization, which also lays a foundation for its clinical further promotion.
This video test provides a white light and dyeing enlarged video data for an endoscopic physician, simulating a clinical environment, providing a full basis for the diagnosis of endoscopic physician, which is more conforming to the trueity of the results. From the front and rear control results of the inner mirror, the results of the inner mirror physicians have the overall theory of AI auxiliary diagnosis, which also illustrates the effectiveness of AI auxiliary diagnosis.
The AI ??auxiliary diagnostic system in this study also has the characteristics of objective and will not fatigue. After the inner mirror doctor, the fatigue problem will occur, which may increase the missed diagnosis. When AI does not have fatigue problems, the physician can refer to the appropriate tips given by Ai in fatigue, thereby increasing the detection rate of gastric cancer underneath. In addition, the judgment of endoscopic physicians may be affected by other factors, and AI has strong objectivity, which gives more objective reference for inner mirrors. Endoscopic physicians will also pay more attention to the AI ??prompts that there is a possible area of ??cancer, thereby reducing the missed diagnosis.
The accuracy of AI is better than that of or without AI-assisted inner mirrors, while the specificity of AI is low compared to endoscopic physicians. For cancer lesions, reduce missed diagnosis, that is, reduced the emergence of false negative cases, and the high sensitivity of our model is just in line with this. As for false positive cases, it can be excluded after the pathology confirmed.
The video verification simulates the real clinical environment and tests the performance of AI in the real-time environment and the performance of the inner mirror physician without AI assist. Ai or AI auxiliary physicians perform a whole than endoscopic physicians. This study demonstrates that the AI ??system has a real-time assistant physician in dyeing enlarged video, which also confirms its good stability and geling. In the future, we will continue to improve models, further in clinical environments to confirm the performance of the AI ??system in dyeing amplification mode.
The reference is omitted.
Chinese Digestive Endoscopy Magazine