Validation of AI-ENDO System in Endoscopic Submucosal Dissection

Gastrointestinal (GI) cancer, including stomach, colorectal, and esophageal cancer, is collectively the most common cancer worldwide and the leading cause of cancer-related death globally. The disease imposes a significant burden in Hong Kong, with more than 6,000 new cases diagnosed annually according to local cancer statistics. Despite substantial advances in understanding its pathology, molecular mechanisms, and therapeutic strategies, the survival rate of advanced GI cancer remains suboptimal, with a 5-year survival rate below 60% for most malignancies.

The introduction of government-led initiatives and international cancer policies has expanded early GI cancer screening programs, enabling more patients to be diagnosed at an earlier stage. Early detection allows for less-invasive and more effective treatment options with improved prognosis and increased curative surgical rates.

Complete tumor resection remains the only curative approach for localized GI cancer. Endoscopic submucosal dissection (ESD), a minimally invasive endoscopic technique, has become an established method for early-stage cancers confined to the mucosal layer. ESD enables en-bloc resection of lesions, achieving lower local recurrence rates compared with conventional endoscopic mucosal resection. It has also demonstrated favorable perioperative outcomes, including shorter operative time, lower complication rates, and reduced hospital stay. In countries where screening programs are well implemented, such as Japan, the number of ESD procedures now exceeds radical gastrectomies, making ESD the preferred treatment modality for early gastric cancer. ESD is also used for the treatment of pre-malignant lesions in the GI tract, preventing progression to invasive cancer.

An ESD procedure involves four main phases: (1) marking of the targeted lesion, (2) submucosal injection to create a protective cushion, (3) circumferential incision, and (4) submucosal dissection and removal of the lesion. Although ESD enables complete resection of lesions regardless of size and has a recurrence rate below 1%, it remains technically demanding and carries risks such as bleeding and perforation. In regions where ESD adoption is lower, the learning curve may exceed several hundred cases before proficiency is achieved. Increasing international guidance and training efforts have encouraged a global expansion in the use of ESD as its clinical benefits become more widely recognized.

Safe and precise dissection of cancerous tissue within the complex gastrointestinal environment is critical. Medical safety events in surgical practice contribute significantly to excess hospital stays, costs, and mortality. ESD is technically challenging, requiring endoscopists to maintain clear visualization of the submucosal layer while preserving a stable endoscopic field. Anatomical complexities such as colon flexures, thin colonic walls, mucosal folds, and peristalsis further increase the difficulty of colorectal ESD and raise the risk of incomplete dissection and perforation. Given these challenges, there is a clear need for advanced technologies that enhance surgical precision and safety through intelligent intraoperative guidance.

To improve surgical quality and patient safety, it is essential to develop systems that learn from expert experience and surgical outcomes, providing real-time insights that help prevent operative errors. Identification of the correct dissection plane, submucosal vessels, and cutting landmarks contributes to safe and complete tumor removal.

An AI-ENDO integrated platform has been developed by the investigators to assist endoscopists in enhancing the safety and efficiency of ESD procedures. The system incorporates multiple intelligent functions:

1. Real-time recognition of ESD surgical phases and performance assessment The AI-ENDO platform analyzes surgical workflow to identify the ongoing procedural phase at each moment. The model was trained on high-quality expert ESD videos collected from Prince of Wales Hospital and externally validated using data from several international centers. The system demonstrated high accuracy, precision, and recall, maintaining consistent performance across datasets. The integrated platform operates with real-time computational efficiency.
2. Real-time identification of safe and dangerous zones during ESD The system performs automated segmentation of intraoperative visual fields, identifying anatomic structures such as mucosa, submucosa, muscle layers, and blood vessels. This enables real-time differentiation between safe dissection zones and areas at higher risk for complications. Validation studies using endoscopic surgical videos demonstrated high segmentation accuracy and low latency, with experiments confirming feasibility during live animal trials. AI assistance was associated with shorter procedural time and reduced bleeding events compared with standard procedures of similar complexity.
3. Intelligent dissection trajectory guidance using reinforcement learning Reinforcement learning and imitation learning methods were incorporated to predict optimal dissection trajectories based on expert surgical demonstrations. The system analyzes surgical context in real time to provide decision support and trajectory prediction. Testing on extensive video datasets demonstrated improved accuracy and generalization compared with existing trajectory prediction frameworks.

Through the integration of these functions, the AI-ENDO multifunctional platform aims to support intraoperative decision-making, reduce adverse events, and shorten the learning curve for ESD procedures.

The purpose of this study is to evaluate the clinical utility and efficacy of the AI-ENDO system in endoscopic submucosal dissection for early gastrointestinal neoplasia. Technical success, clinical success, and post-ESD adverse events will be systematically assessed.