Research On Artificial Intelligence Assisted Endoscopy In Diagnosing Early Esophageal Tumor

BackgroundEsophageal cancer is one of the most common malignant tumor worldwide,with high morbidity and mortality.In China,the incidence of esophageal cancer has declined in recent years.But its mortality is still relatively high and ranked fourth.Esophageal cancer is a major disease that threatens national health.Most patients with esophageal cancer are already in the advanced stage when they are diagnosed.Even after surgery,radiotherapy and chemotherapy,these patients hold a 5-year survival rate less than 20%.However,if it can be early diagnosis by screening and be treated activity,the 5-year survival rate of early esophageal cancer can reach more than 90%.Therefore,early diagnosis and treatment are the key to manage esophageal cancer.Due to various factors,the detection rate of early esophageal cancer is unsatisfactory and limits the therapeutic effect of esophageal cancer in China.With the development of technology,artificial intelligence is widely used in the field of medical imaging.It is expected that this technology might also greatly increase the detection rate of early esophageal cancer.AimEstablish an artificial intelligence-assisted endoscopic diagnosis system for early esophageal cancer and precancerous lesions,aiming to improve the detection rate with white light endoscope.By improving the algorithm,make the model have the ability to determine lesion boundaries and real-time application.Methods480 cases of early esophageal cancer and precancerous lesions was retrospectively selected,which were pathologically confirmed as early esophageal cancer after EMR or ESD.A total of 3328 endoscopic images with the lesions were selected among these patients.The lesions of selected images were marked by two experienced endoscopic physicians using a closed curve.Endoscopic images with iodine staining of the lesion were review to determine the boundary.2036 of these images were used as the training set,the others were used as the test set.At the same time,1050 normal endoscopy images of esophageal were selected to be added to the test set as a negative control.The classification model of early esophageal cancer and precancerous lesions was established by YOLOv3 algorithm;The lesion boundary recognition model was established by Bisenet algorithm.The ROC curve was drawn to get the area under the curve(AUC value).Calculate the overlapping area of the lesion range recognized by the lesion boundary recognition model and the actual lesion range.The speed of picture recognition is recorded to test whether it meets the requirements of real-time identification.ResultsThis diagnostic system has high efficiency for the detection of early esophageal cancer and precancerous lesions: when the diagnostic threshold is set to 70%,the correct rate is 99.06%,the sensitivity is 98.76%,the specificity is 99.43%,the positive predictive value is 99.53 %,and negative predictive value is 98.49%.The area under the curve(AUC value)was 0.928.The diagnosis system has a high accuracy of identifying the lesion area.The overlapping area fluctuates between 65%-95%.The recognition speed of this diagnostic system can realize real-time recognition.ConclusionThe artificial intelligence-assisted model has a high efficiency to detect early esophageal cancer and precancerous lesions and identify boundary of lesion.Through model improvement,real-time application can be achieved.