LAUSR

Early detection of pharyngeal cancer contributes to providing excellent long-term outcomes and preserving pharyngeal function. Thus, an efficient detection system for superficial pharyngeal cancer (SPC) is necessary worldwide, especially in Asia. However, early detection proves challenging because SPC is typically flat with subtle color changes. Although image enhanced endoscopy, such as narrow band imaging (NBI), substantially helps us detect SPC, it is still difficult even for experienced endoscopists. Based on this background, artificial intelligence (AI) technologyhasbeen recently utilized for the earlydetection ofpharyngeal cancer.Mascharak et al.first reported imageprocessing andbasic machine learning techniques to automate the assessment of oropharyngeal cancer using white light endoscopy (WLE) and NBI. However, the sensitivity and specificitywere unsatisfactory, even using NBI, perhaps owing to a sample size of only 30 patients. Later, Tamashiro et al. developed an AI system using 5403 still images of pharyngeal cancer. In this study, as expected, the AI system took only 28 seconds to analyze 1912 validation still images consisting of 928 and 984 with or without pharyngeal cancers, respectively. The sensitivity and the specificity per image analysis were 79.7% and were 57.1%, respectively. This AI system demonstrated sensitive pharyngeal cancer detection. However, validation using still images would be subject to bias because still images are usually taken under good conditions (e.g., adequate angle, adequate distance, and in focus). Although the AI system for still image diagnosis can be utilized for doublechecking after screening endoscopy, it is unsuitable for real-world endoscopic diagnosis. In this issue of Digestive Endoscopy, Kono et al. present the diagnostic performance of their AI system for pharyngeal cancer detection using an independent validation dataset of 25 videos of pharyngeal cancer and 36 videos of nonpharyngeal cancer. The sensitivity, specificity, and accuracy for detecting cancer were 92%, 47%, and 66%, respectively. This study highlighted that the AI system allowed for sensitive and real-time endoscopic pharyngeal cancer detection. The validation with video images is more challenging than that with still images because the images inevitably include poor-quality features (e.g., defocus, light reflection, mucus, or saliva), which were excluded in Tamashiro et al. Nevertheless, the video-based validation provides a more realistic and practical assessment of the diagnostic performance of the AI system. The concept of this study is highly valuable because real-time pharyngeal cancer detection is more supportive and educational than still image assessment, particularly for inexperienced endoscopists. Moreover, the concept could play a role in quality assurance. As shown above, the AI system was a promising tool for pharyngeal cancer detection. However, we want to discuss areas of further improvement of the AI system because it appears to be at an investigational stage that requires adaptation for the real-world endoscopic detection. First, further robust data of its sensitivity should be accumulated. In terms of imaging modality, the sensitivity of the AI system appears to be higher for NBI thanWLE, as it is in human diagnosis. Although the sensitivity was reported to be relatively favorable, the sample size was small for its use in real clinical practice. It is desirable to further train the AI system with more pharyngeal cancer images, primarily SPC. Close collaboration with otolaryngologists is considered necessary to increase the sample size as they still manage most SPC. Also, given the low prevalence of pharyngeal cancer, it is not realistic to establish the training in a single center, even if endoscopists focus on high-risk populations. A specific multicenter database such as the Japan Endoscopy Database (JED) will help collect enough samples of rare diseases and establish a better trained AI system. Secondly, more importantly, specificity should also be improved. In both studies, the specificity was unsatisfactory, as the authors pointed out. Consequently, the positive predictive value was low. The positive predictive value of the AI system will become further lower in real clinical practice given the low prevalence of pharyngeal cancer and the proportion of pharyngeal cancer and non-neoplastic lesions. It means that endoscopists will face excessive caution introduced by the AI system during an endoscopy examination. The main reason to explain the low specificity was insufficient training of noncancerous lesions, due to having included only images of pharyngeal cancer in the training set. However, a higher specificity will not be achieved by simply educating with non-