Introduction
Endoscopic Submucosal Dissection (ESD) has emerged as a vital, minimally invasive surgical technique for the treatment of early-stage gastrointestinal tumors, such as Early Esophageal Cancer (EEC). A successful ESD procedure relies heavily on the "marking" phase, where an endoscopist identifies the lesion's boundary and places physical marks at a safe distance outside it. These marks serve as a roadmap for the subsequent circumferential incision.
Despite its effectiveness, ESD is technically demanding and highly subjective. The decision-making process regarding where to place marks and where to cut varies significantly between practitioners based on their individual experience. This variance leads to inconsistencies in surgical quality: a margin that is too narrow risks leaving malignant tissue behind, while a margin that is too wide increases the risk of procedural complications. To address this lack of standardization, researchers have developed an intelligent automatic system designed to provide real-time guidance and quantify the marking and incision process.
The AI Assistant System
The proposed AI system acts as a real-time "digital ruler" for the surgeon, integrating two primary functional modules to guide the procedure from boundary identification to the final incision.
- Marking Points Suggestion Module: Utilizing computer vision, this module detects the lesion boundary in real-time. It automatically calculates and overlays suggested marking positions at a precise, adjustable distance—typically 0.3 cm—outside the identified boundary. This ensures that the initial safety margin is consistent and evidence-based.
- Cutting Line Suggestion Module: After the endoscopist places physical marks on the mucosa, the system detects these marks and generates a suggested mucosal cutting path. This provides a visual guide for the circumferential incision, ensuring the surgeon follows the intended path accurately.
Methodology and Datasets
The system was trained and evaluated using two distinct datasets curated from clinical cases of Early Esophageal Cancer (EEC):
- Lesion Boundary Dataset: This included 499 endoscopic images from 20 cases. These images utilized Lugol’s staining, a clinical technique that highlights cancerous cells, to provide the AI with clear ground-truth data for lesion segmentation.
- ESD Procedure Dataset: This comprised 3,397 detailed images captured during actual ESD procedures across 12 cases. These images were used to train the system to recognize the artificial marks made by surgeons during live operations.
By combining these datasets, the researchers built a robust algorithm capable of synchronizing with the surgeon's workflow, moving from lesion detection to mark recognition and finally to incision guidance.
Key Results
The experimental results demonstrated that the AI system achieves high technical proficiency in the tasks required for surgical guidance:
- Lesion Segmentation: The system achieved a Dice score of 96.1%, indicating an exceptionally high level of accuracy in identifying the boundaries of early esophageal lesions.
- Mark Detection: The system reached an Average Precision (AP50) of 72.1% in detecting the physical marks placed by the endoscopist during the procedure.
- Functional Integration: Beyond these metrics, the system successfully demonstrated its ability to display real-time guidance for both the marking phase (using the 0.3 cm boundary offset) and the subsequent incision phase.
Discussion and Clinical Impact
The core innovation of this research lies in shifting the role of AI from a diagnostic tool to a procedural assistant. By quantifying what was previously a subjective decision-making process, the system helps "democratize" surgical precision. This is particularly relevant for standardizing ESD procedures across different levels of healthcare facilities, potentially allowing less experienced endoscopists in smaller hospitals to achieve the same consistency and safety margins as experts in high-volume centers.
Furthermore, the system creates a closed feedback loop: it suggests where to mark, detects the marks once they are placed, and then calculates the final cutting path. This ensures the AI remains synchronized with the surgeon's physical actions throughout the operation.
Conclusion and Future Directions
While the technical performance of the system is promising, the study was limited by a relatively small cohort of 32 cases. Future work will require larger-scale, multi-center trials to validate the system across different endoscopic equipment and a wider variety of lesion appearances. Additionally, further research is needed to determine if the use of this AI assistant directly translates to improved clinical outcomes, such as reduced procedure times, lower complication rates, or higher R0 resection rates.
Ultimately, this automatic system represents a significant step toward the standardization of ESD, offering a path toward more consistent and reliable surgical care for patients with early-stage gastrointestinal cancers.
Summary updated