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

Preventing intraoperative bleeding in endoscopic submucosal dissection (ESD) relies on the timely identification and pre-coagulation of submucosal vessels. We aim to reduce bleeding risk by delivering accurate vessel segmentation with uncertainty awareness.

We present ESD-VesNet, a prompt-free SAM3-based framework with an evidential head that outputs vessel probability and pixel-wise uncertainty. The model is trained in a false-positive-aware manner using our developed ESD-Vessel dataset, which includes 2401 positives (images with clear vessel annotation) and 708 curated hard negatives (images may be confused with vessels), combined with online uncertainty-guided hard negative mining. The model is trained in a false-positive-aware manner using the ESD-Vessel dataset, which includes 2401 positive samples (images with clear vessel annotations) and 708 curated hard negatives (images visually confusable with vessels), combined with online uncertainty-guided hard negative mining. By suppressing false positive predictions, our approach has the potential to minimize unnecessary interventions, such as unwarranted coagulation.

On the validation set, ESD-VesNet achieves Dice $_{\mathrm{FG}}^{}$ 52.11%, IoU $_{\mathrm{FG}}^{}$ 35.23%, mDice 75.70%, and Vessel Detection Rate 96.81%, while keeping Background False Positive Rate at 0.61%. Structural quality is high (E-measure 0.7275; S-measure 0.7223). Qualitative analyses show robustness under bubbles, debris, reflections, and blood seepage.

Integrating SAM3 with evidential uncertainty and hard negative mining yields a clinically oriented vessel segmentation system that maintains high detection sensitivity with low false positives, supporting safer ESD.