Automated artificial intelligence–based phase-recognition system for esophageal endoscopic submucosal dissection (with video)

Background and Aims Endoscopic submucosal dissection (ESD) for superficial esophageal cancer is a multistep treatment involving several endoscopic processes. Although analyzing each phase separately is worthwhile, it is not realistic in practice owing to the need for considerable manpower. To solve this problem, we aimed to establish a state-of-the-art artificial intelligence (AI)–based system, specifically, an automated phase-recognition system that can automatically identify each endoscopic phase based on video images. Methods Ninety-four videos of ESD procedures for superficial esophageal cancer were evaluated in this single-center study. A deep neural network–based phase-recognition system was developed in an automated manner to recognize each of the endoscopic phases. The system was trained with the use of videos that were annotated and verified by 2 GI endoscopists. Results The overall accuracy of the AI model for automated phase recognition was 90%, and the average precision, recall, and F value rates were 91%, 90%, and 90%, respectively. Two representative ESD videos predicted by the model indicated the usability of AI in clinical practice. Conclusions We demonstrated that an AI-based automated phase-recognition system for esophageal ESD can be established with high accuracy. To the best of our knowledge, this is the first report on automated recognition of ESD treatment phases. Because this system enabled a detailed analysis of phases, collecting large volumes of data in the future may help to identify quality indicators for treatment techniques and uncover unmet medical needs that necessitate the creation of new treatment methods and devices. Endoscopic submucosal dissection (ESD) for superficial esophageal cancer is a multistep treatment involving several endoscopic processes. Although analyzing each phase separately is worthwhile, it is not realistic in practice owing to the need for considerable manpower. To solve this problem, we aimed to establish a state-of-the-art artificial intelligence (AI)–based system, specifically, an automated phase-recognition system that can automatically identify each endoscopic phase based on video images. Ninety-four videos of ESD procedures for superficial esophageal cancer were evaluated in this single-center study. A deep neural network–based phase-recognition system was developed in an automated manner to recognize each of the endoscopic phases. The system was trained with the use of videos that were annotated and verified by 2 GI endoscopists. The overall accuracy of the AI model for automated phase recognition was 90%, and the average precision, recall, and F value rates were 91%, 90%, and 90%, respectively. Two representative ESD videos predicted by the model indicated the usability of AI in clinical practice. We demonstrated that an AI-based automated phase-recognition system for esophageal ESD can be established with high accuracy. To the best of our knowledge, this is the first report on automated recognition of ESD treatment phases. Because this system enabled a detailed analysis of phases, collecting large volumes of data in the future may help to identify quality indicators for treatment techniques and uncover unmet medical needs that necessitate the creation of new treatment methods and devices.