Establishment of an artificial intelligence assisted diagnosis model based on deep learning for recognizing gastric lesions and their locations under gastroscopy in real time
- VernacularTitle:基于深度学习的人工智能辅助胃镜下实时识别病变及位置模型的建立
- Author:
Xian GUO
1
;
Ying-Yang WU
;
Ai-Rui JIANG
;
Chao-Qiang FAN
;
Xue PENG
;
Xu-Biao NIE
;
Hui LIN
;
Jian-Ying BAI
Author Information
- Keywords: artificial intelligence; gastric diseases; early gastric cancer; deep learning; gastroscopy examination
- From: Journal of Regional Anatomy and Operative Surgery 2024;33(10):849-854
- CountryChina
- Language:Chinese
- Abstract: Objective To construct an artificial intelligence assisted diagnosis model based on deep learning for dynamically recognizing gastric lesions and their locations under gastroscopy in real time,and to evaluate its ability to detect and recognize gastric lesions and their locations.Methods The gastroscopy videos of 104 patients in our hospital was retrospectively analyzed,and the video frames were manually annotated.The annotated picture frames of lesion category were divided into the training set and the validation set according to the ratio of 8∶2,and the annotated picture frames of location category were divided into the training set and the validation set according to the patient sources at the ratio of 8∶2.These sets were utilized for training and validating the respective models.YoloV4 model was used for the training of lesion recognition,and ResNet152 model was used for the training of location recognition.The accuracy,sensitivity,specificity,positive predictive value,negative predictive value and location recognition accuracy of the auxiliary diagnostic model were evaluated.Results A total of 68 351 image frames were annotated,with 54 872 frames used as the training set,including 41 692 frames for lesion categories and 13 180 frames for location categories.The validation set consisted of 13 479 frames,comprising 10 422 frames for lesion categories and 3 057 frames for location categories.The lesion recognition model achieved an overall accuracy of 98.8%,with a sensitivity of 96.6%,specificity of 99.3%,positive predictive value of 96.3%,and negative predictive value of 99.3% in validation set.Meanwhile,the location recognition model demonstrated an top-5 accuracy of 87.1% .Conclusion The artificial intelligence assisted diagnosis model based on deep learning for real-time dynamic recognition of gastric lesions and their locations under gastroscopy has good ability in lesion detection and location recognition,and has great clinical application prospects.
