[Paper Review] Cholec80-port: A Geometrically Consistent Trocar Port Segmentation Dataset for Robust Surgical Scene Understanding

Shunsuke Kikuchi, Akimasa Kouno+1 more|arXiv (Cornell University)|Feb 19, 2026

Surgical Simulation and Training0 citations

TL;DR

Introduces Cholec80-port, a geometrically consistent trocar port segmentation dataset derived from Cholec80, with a sleeve-based annotation SOP excluding the central lumen, plus cleansing of existing datasets and a baseline model. Shows improved cross-dataset robustness when trained on the new dataset.

ABSTRACT

Trocar ports are camera-fixed, pseudo-static structures that can persistently occlude laparoscopic views and attract disproportionate feature points due to specular, textured surfaces. This makes ports particularly detrimental to geometry-based downstream pipelines such as image stitching, 3D reconstruction, and visual SLAM, where dynamic or non-anatomical outliers degrade alignment and tracking stability. Despite this practical importance, explicit port labels are rare in public surgical datasets, and existing annotations often violate geometric consistency by masking the central lumen (opening), even when anatomical regions are visible through it. We present Cholec80-port, a high-fidelity trocar port segmentation dataset derived from Cholec80, together with a rigorous standard operating procedure (SOP) that defines a port-sleeve mask excluding the central opening. We additionally cleanse and unify existing public datasets under the same SOP. Experiments demonstrate that geometrically consistent annotations substantially improve cross-dataset robustness beyond what dataset size alone provides.

Motivation & Objective

• Highlight the need for geometry-faithful port labeling to improve segmentation and downstream geometry tasks in laparoscopic surgery.
• Present a large-scale, sleeve-consistent port segmentation dataset derived from Cholec80.
• Provide data cleansing of existing datasets to align with the new annotation SOP.
• Release a pretrained port segmentation model and tooling for downstream CV research.

Proposed method

• Define port-sleeve as the rigid sleeve visible beyond the internal valve, excluding the central lumen.
• Annotate 38,434 frames from Cholec80 using CVAT with temporal verification for boundary fidelity.
• Cleansing and unification of m2caiSeg and GynSurg under the sleeve-consistent SOP.
• Train a ConvNeXt-Base encoder with a U-Net decoder using Dice plus BCE loss.
• Evaluate with Dice score on port-present frames and frame-level Detect F1 across datasets.

Experimental results

Research questions

• RQ1 Does geometrically consistent sleeve-based port labeling improve segmentation robustness and generalization across laparoscopic datasets?
• RQ2 Can a dedicated port annotation SOP reduce label noise and improve cross-dataset transfer for port segmentation?
• RQ3 How does a model trained on Cholec80-port perform on in-domain and cross-domain port segmentation tasks?
• RQ4 What is the impact of cleansing existing datasets on transfer performance?
• RQ5 What are the main failure modes under the sleeve-consistent definition?

Key findings

• Models trained on Cholec80-port achieve strong in-domain Dice 0.862 and Detect F1 0.856 on the Cholec80-port test set.
• Cholec80-port training outperforms m2caiSeg-trained models on the m2caiSeg test set, indicating robustness beyond the immediate dataset.
• Cross-dataset generalization to GynSurg remains challenging due to domain shift in port appearance and lighting.
• Cleansing and reformulating m2caiSeg and GynSurg under the sleeve SOP substantially improves transfer robustness, while unsanctioned labels hinder performance.
• A combined cleaned dataset yields notable gains (Dice 0.812–0.823 Detect F1) but domain shift persists, highlighting the need for more visual diversity.