ESD-VesNet: uncertainty-aware vessel segmentation network for endoscopic submucosal dissection with hard negative...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Endoscopy
• Surgical Safety

Background:

• Intraoperative bleeding during endoscopic submucosal dissection (ESD) is a significant risk.
• Accurate identification and pre-coagulation of submucosal vessels are crucial for preventing bleeding.
• Current methods may lack precision in vessel detection, leading to potential complications.

Purpose of the Study:

• To develop an AI-driven system for precise submucosal vessel segmentation in ESD.
• To enhance surgical safety by reducing the risk of intraoperative bleeding.
• To provide uncertainty awareness in vessel detection to minimize unnecessary interventions.

Main Methods:

• Introduction of ESD-VesNet, a framework based on SAM3 with an evidential head for probability and uncertainty output.
• Training on a custom ESD-Vessel dataset with positive and hard negative samples.
• Implementation of false-positive-aware training and uncertainty-guided hard negative mining.

Main Results:

• ESD-VesNet achieved a Vessel Detection Rate of 96.81% with a low Background False Positive Rate of 0.61%.
• High structural quality metrics (E-measure 0.7275, S-measure 0.7223) were reported.
• The model demonstrated robustness in challenging conditions like bubbles, debris, and blood seepage.

Conclusions:

• The integration of SAM3, evidential uncertainty, and hard negative mining creates a clinically valuable vessel segmentation system.
• ESD-VesNet offers high detection sensitivity and low false positives, contributing to safer ESD procedures.
• This approach supports clinicians by providing reliable vessel identification, minimizing risks.