Background and study aims
Completeness of esophagagogastroduodenoscopy (EGD) varies among endoscopists, leading to a high miss rate for gastric neoplasms. This study aimed to determine the effect of the Cerebro real-time artificial intelligence (AI) system on completeness of EGD for endoscopists in early stages of training.
Patients and methods
The AI system was built with CNN and Motion Adaptive Temporal Feature Aggregation (MA-TFA). A prospective sequential cohort study was conducted. Endoscopists were taught about the standardized EGD protocol to examine 27 sites. Then, each subject performed diagnostic EGDs per protocol (control arm). After completion of the required sample size, subjects performed diagnostic EGDs with assistance of the AI (study arm). The primary outcome was the rate of completeness of EGD. Secondary outcomes included overall inspection time, individual site inspection time, completeness of photodocumentation, and rate of positive pathologies.
Results
A total of 466 EGDs were performed with 233 in each group. Use of AI significantly improved completeness of EGD [mean (SD) (92.6% (6.2%) vs 71.2% (16.8%)]; P <0.001 (95% confidence interval 19.2%–23.8%, SD 0.012). There was no difference in overall mean (SD) inspection time [765.5 (338.4) seconds vs 740.4 (266.2); P=0.374]. Mean (SD) number of photos for photo-documentation significantly increased in the AI group [26.9 (0.4) vs 10.3 (4.4); P <0.001]. There was no difference in detection rates for pathologies in the two groups [8/233 (3.43%) vs 5/233 (2.16%), P=0.399].
Conclusions
Completeness of EGD examination and photodocumentation by endoscopists in early stages of are improved by the AI-assisted software Cerebro.
Background and study aims: Endoscopic submucosal dissection (ESD) is a challenging minimally invasive resection technique with a long training period and relevant operator-dependent complications. Real-time artificial intelligence (AI) orientation support may improve safety and intervention speed.
Methods: A total of 1011 endoscopic still images from 30 ESDs were annotated for relevant anatomical structures and used for training of a deep learning algorithm. After internal and external validation, this algorithm was applied to 12 ESDs performed by either one expert or one novice in ESD using an in vivo porcine model.
Results: External validation yielded mean Dice Scores of 88%, 60%, 58%, and 92% for background, submucosal layer, submucosal blood vessels, and muscle layer, respectively. The system was successfully applied during all 12 ESDs. All resections were completed en bloc and without complications.
Conclusions: In this proof-of-concept study, feasibility of a real-time AI algorithm for anatomical structure delineation and orientation support during ESD was evaluated. The application proved safe and appropriate for routine procedures in humans. Further studies are needed to elucidate a potential clinical benefit of this new technology.
