TL;DR
This study introduces and validates a novel AI-powered lesion labeling system designed to assist endoscopists in performing endoscopic submucosal dissection (ESD) for esophageal lesions. The system demonstrated comparable efficacy to conventional ESD in high-volume centers, particularly in improving complete resection rates for beginners in low-volume settings, suggesting a path to democratize advanced endoscopic procedures.
What Problem Does This Paper Address?
Esophageal cancer is a significant global health concern, and endoscopic submucosal dissection (ESD) is the preferred treatment for early-stage lesions due to its minimally invasive nature and ability to achieve curative resection. A critical factor for successful ESD is the accurate delineation of lesion margins to ensure complete removal and prevent recurrence. However, this task is highly subjective and relies heavily on the endoscopist's experience. In low-volume centers, where expertise may be limited, achieving high rates of complete resection (specifically, negative lateral margins) is challenging, often leading to suboptimal outcomes. This paper addresses the need for an objective, AI-driven tool to assist endoscopists, especially those with less experience in lower-volume settings, in precisely identifying and marking lesion boundaries during ESD.
Methods
The researchers developed and implemented a real-time AI lesion labeling system that integrates with endoscopic imaging (specifically Narrow Band Imaging - NBI). This system utilizes a deep learning architecture, building upon a VGG-16 backbone, to detect and classify esophageal lesions. The core innovation lies in its ability to output lesion segmentation maps, which clinicians then use to strategically identify surrounding mucosa at the lesion boundary. This optimized margin recognition model was then applied in a prospective cohort study.
The study involved two groups:
- AI-assisted ESD group: Patients treated in a low-volume center equipped with the AI lesion labeling system, with ESD performed primarily by beginners.
- Conventional ESD group: Patients treated in a high-volume center with experienced endoscopists, undergoing conventional ESD without AI assistance.
The primary outcome measured was the complete lateral resection rate, with secondary outcomes including en-bloc resection rate, procedure duration, specimen diameter, and complication rates.
Datasets
The study utilized patient data collected prospectively from January 2021 to December 2022 across two endoscopy centers in Shanghai, China. A total of 402 lesions from 355 patients were included after excluding those with pathological outcomes other than esophageal squamous cell carcinoma or precancerous lesions. The AI system's architecture was previously validated on large-scale colon and esophageal datasets, and its components were trained using the ImageNet dataset.
Key Insights and Ideas
- AI as an Experience Equalizer: The AI system appears to bridge the experience gap for endoscopists in low-volume centers. Beginners using the AI system achieved outcomes comparable to experienced endoscopists in high-volume centers, particularly regarding complete lateral resection rates.
- Beyond Detection: Surgical Assistance: This study moves beyond AI's role in lesion detection and diagnosis to its application in guiding surgical procedures (ESD), specifically in marking resection margins.
- Tolerable Trade-off: While the AI-assisted ESD group experienced longer procedure times, this was deemed a tolerable drawback given the significant improvement in achieving complete resection, especially in resource-constrained settings.
- Safety Profile: The AI-assisted ESD procedure demonstrated a good safety profile, with low rates of adverse events like delayed bleeding, thoracic effusion, and perforation.
Key Results
- Complete Lateral Resection: The AI-assisted ESD group achieved a complete lateral resection rate of 90.0% (180/200), which was non-inferior to the conventional ESD group's rate of 92.1% (186/202) (P = 0.465). This is notably higher than previously reported rates for low-volume centers.
- En-bloc Resection: Both groups showed high en-bloc resection rates: 98.5% (197/200) for the AI group and 98.5% (199/202) for the conventional group.
- Procedure Duration: The AI-assisted ESD group had a significantly longer median procedure duration (82 minutes) compared to the conventional ESD group (49 minutes) (P < 0.001).
- Adverse Events: Delayed bleeding occurred in 0.5% of AI-assisted cases, minor thoracic effusion in 1.0%, and perforation in 1.0%.
Limitations and Future Work
The authors acknowledge several limitations:
- Non-Randomized Design: The study was prospective but not randomized. The authors suggest this was an exploratory investigation to assess feasibility and efficacy, paving the way for future randomized controlled trials.
- Potential for Operator Bias: The AI system might have made operators more concentrated, potentially influencing outcomes. Future studies could benefit from a sham system control group to blind participants.
- Prolonged Procedure Time: While deemed tolerable, the increased procedure time in the AI group warrants further investigation and optimization.
Future work could focus on larger, randomized trials to further validate these findings and explore the broader applicability of AI in complex endoscopic procedures across various settings. The authors also highlight the potential cost-effectiveness of such AI systems in preventing costly recurrences and improving healthcare equity.
Summary updated