Deep learning-based phase recognition system for endoscopic submucosal dissection of early gastrointestinal cancer
Background and Aims: Endoscopic submucosal dissection (ESD) has become a standard treatment for early-stage gastrointestinal (GI) cancer. Automated recognition for various ESD (endoscopic submucosal dissection) phase helps in ESD training and execution. The existing phase recognition systems are lim...
Saved in:
| Main Authors: | , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-02-01
|
| Series: | The Lancet Regional Health. Western Pacific |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2666606524004395 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Background and Aims: Endoscopic submucosal dissection (ESD) has become a standard treatment for early-stage gastrointestinal (GI) cancer. Automated recognition for various ESD (endoscopic submucosal dissection) phase helps in ESD training and execution. The existing phase recognition systems are limited by their scope, focusing on specific types of ESD and lacking comprehensive phase identification. This study aims to develop a deep learning-based phase recognition system, named AI-ESD, that can accurately identify and categorize all phases of ESD for various gastrointestinal locations. Methods: Using 234 ESD videos from a single center, we trained AI-ESD to recognize 8 phases: preparation, estimation, marking, submucosal injection, incision, submucosal dissection, vessel treatment, and clips. Performance was evaluated on an independent validation set (n=69) and videos from novice endoscopists (n=34) using accuracy, sensitivity, specificity, F1-score and area under the receiver operating characteristic curve (AUROC). Subgroup analyses were performed by tumor location, endoscopist, and lesion size. Findings: The AUROC, accuracy, and specificity for all ESD phases was higher than 0.9 with sensitivity ranging from 0.880 (phase of incision) to 0.963 (phase of clips) and F1 score ranging from 0.823 (phase of incision) to 0.948 (phase of marking). The macro-average AUROC, accuracy, sensitivity, specificity, and F1 score were 0.987, 0.975, 0.919, 0.985, and 0.908, respectively. The AI-ESD model demonstrated nearly similar results in the subgroup analysis to overall analysis across all phases in the validation cohort. Similar results for novice cohort were obtained with macro-average AUROC, accuracy, sensitivity, specificity, and F1 score being 0.981, 0.971, 0.901, 0.981, and 0.897, respectively. Sensitivity was lowest for the preparation phase (0.881 validation, 0.848 novice). Interpretation: We developed AI-ESD for highly accurate automated phase recognition system for full gastrointestinal ESD, including gastric, esophageal, and colorectal ESD. Further external validation in diverse settings is warranted. |
|---|---|
| ISSN: | 2666-6065 |