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

This paper introduces Cholec80-port, a high-fidelity dataset and rigorous annotation standard for trocar ports that excludes central lumens to ensure geometric consistency, thereby significantly enhancing the robustness of downstream surgical scene understanding tasks like 3D reconstruction and visual SLAM.

The Gist

Imagine you are trying to take a panoramic photo of a beautiful landscape, but every time you move the camera, a giant, shiny metal pole (the tripod leg) keeps getting in the way. Not only does it block your view, but its shiny surface reflects the sky and trees, creating confusing "ghost" images that trick your camera's software into thinking the scenery is moving when it's actually the pole.

This is exactly the problem surgeons and robots face inside the human body during laparoscopic (keyhole) surgery, and this paper introduces a new solution called Cholec80-port.

Here is the breakdown in simple terms:

1. The Problem: The "Shiny Gatekeeper"

In keyhole surgery, doctors insert a camera through a small tube called a trocar port (like a gatekeeper) that goes through the belly wall.

• The Issue: Sometimes this port stays in the camera's view. It is made of shiny metal or plastic and has a hole in the middle.
• Why it's bad for robots: Computers trying to build a 3D map of the inside of the body (like a GPS for surgery) get confused by this port. Because the port is shiny and doesn't move (unlike the liver or heart), the computer thinks the entire world is moving around it. It's like trying to navigate a car while staring at a static, shiny sticker on your windshield; the computer gets lost.

2. The Missing Map

To teach computers to ignore this port, researchers need a "map" (a dataset) where the port is clearly labeled.

• The Old Maps: Previous maps existed, but they were flawed. Some were too small. Others had a weird rule where they labeled the hole in the port as part of the port. This is like drawing a circle around a donut and including the empty space in the middle as "donut." This confuses the computer because it can see the organs through that hole.
• The Result: Computers trained on these old maps kept making mistakes.

3. The New Solution: Cholec80-port

The authors created a new, high-quality dataset called Cholec80-port. Think of this as a "Gold Standard" training manual for robots.

How they fixed the "Donut" problem:
They introduced a new rule called Geometric Consistency.

• Old Way: Label the whole port, including the hole.
• New Way (The "Sleeve" Rule): Only label the solid metal ring (the sleeve). Leave the hole empty so the computer knows, "Ah, I can see the organs through this hole; don't treat the hole as part of the obstacle."

They also went through old datasets and "cleaned" them, fixing the bad labels just like a teacher correcting a student's homework before a test.

4. The Training Camp

They took 20 hours of real surgery videos and carefully labeled over 38,000 frames. They taught a computer model (a digital brain) to recognize these "port sleeves" using the new, strict rules.

The Results:

• Better Vision: When tested, the new model was much better at spotting the port and ignoring the hole.
• Super Power: Even when they tested this new model on old, messy data it had never seen before, it still performed better than models trained on the old, messy data. It's like a student who learned the principles of math doing better on a new test than a student who just memorized the answers to the old test.

5. Why This Matters

This isn't just about drawing lines on a picture. By teaching computers to ignore these "shiny gatekeepers," we can:

• Build better 3D maps of the inside of the body.
• Stitch video frames together smoothly without glitches.
• Help surgical robots navigate more safely without getting confused by the equipment they are holding.

In a nutshell: The authors built a better "training school" for surgical robots, teaching them to distinguish between the "gate" (the port) and the "view through the gate" (the organs), so the robots don't get lost in their own reflection.

Technical Summary

1. Problem Statement

In laparoscopic surgery, precise segmentation of instruments and anatomy is critical for downstream geometric tasks such as image stitching, 3D reconstruction, and Visual SLAM (vSLAM). A specific challenge arises from trocar ports (the physical gateway through the abdominal wall).

• Geometric Interference: Trocar ports are often specular, textured, and camera-fixed. They attract excessive feature points and introduce strong, non-anatomical features that bias matching algorithms, increase geometric error, and cause alignment artifacts.
• Data Scarcity & Inconsistency: While ports are detrimental to geometry-based methods, explicit labels are missing from most large-scale datasets due to de-identification concerns (ports reveal the external environment).
• Limitations of Existing Datasets:
  • m2caiSeg: Too small ($N=370$) and suffers from annotation artifacts (interpolation noise).
  • GynSurg: Larger ($N=4873$) but uses a "hole-filling" policy where the central lumen (opening) is masked. This is geometrically inconsistent because it obscures valid anatomical pixels visible through the port opening, disrupting feature extraction.

2. Methodology

2.1 Dataset Acquisition and Sampling

• Source: The authors utilized the first 20 videos from the Cholec80 dataset.
• Sampling Strategy: To balance diversity and efficiency, they sampled every 30th frame, resulting in 38,434 annotated frames.
• Split: The data was split at the video level to prevent leakage:
  • Training: Videos 01–08
  • Validation: Videos 09–10
  • Testing: Videos 11–20
• Positive Samples: 1,398 frames contain visible ports, significantly larger than the positive samples in m2caiSeg (255) and GynSurg (130).

2.2 Annotation SOP: Geometric Consistency

The core innovation is a rigorous Standard Operating Procedure (SOP) for annotation that prioritizes geometric fidelity:

• Definition: The target is defined strictly as the port-sleeve (the rigid metallic/plastic cylinder visible beyond the internal valve).
• Exclusion: The central lumen (opening) is explicitly excluded from the mask.
  • Reasoning: Masking the lumen suppresses anatomically valid pixels visible through the hole, creating geometric inconsistencies for image-plane aggregation and feature extraction.
• Handling Ambiguity: For frames with unclear boundaries due to specular highlights or orientation, annotators used temporal context (neighboring frames) to verify the physical extent of the sleeve, distinguishing it from transient reflections.

2.3 Data Cleansing and Unification

The authors applied the new SOP to cleanse and unify existing datasets:

• m2caiSeg: Re-annotated to remove interpolation artifacts. Only a small subset met the high-quality sleeve-consistent criteria.
• GynSurg: Corrected the "hole-filling" issue by segmenting the central lumen and subtracting it from the original polygons to derive sleeve-only masks.

2.4 Model and Training Protocol

• Architecture: ConvNeXt-Base encoder with a U-Net decoder for binary semantic segmentation.
• Loss Function: Combined Dice Loss and Binary Cross-Entropy (BCE).
• Hyperparameters: AdamW optimizer, learning rate $5 \times 10^{-5}$, batch size 16, input resolution $384 \times 384$.

2.5 Evaluation Metrics

Two complementary metrics were used:

1. Dice Score: Computed only on frames where ports are present ($GT > 0$) to measure boundary recovery fidelity.
2. Detect F1: A frame-level metric classifying a frame as positive if at least one port pixel is predicted, measuring robustness in detection.

3. Key Results

• In-Domain Performance: Training on Cholec80-port achieved a Dice score of 0.862 and Detect F1 of 0.856 on the Cholec80-port test split.
• Cross-Dataset Generalization:
  • The Cholec80-port model outperformed the m2caiSeg-trained model even when evaluated on the m2caiSeg test set (Dice 0.488 vs. 0.448), suggesting that geometrically consistent labels improve robustness beyond specific dataset fitting.
  • Generalization to GynSurg remains challenging (Dice 0.611) due to domain shifts in materials and lighting, but the Combined (cleaned) dataset significantly improved performance across all splits (e.g., GynSurg test Dice improved to 0.819).
• Impact of Cleansing: Ablation studies confirmed that rigorous SOP-level geometric consistency is a dominant factor for cross-dataset robustness. Models trained on uncleaned data showed substantially worse transfer performance.

4. Key Contributions

1. Rigorous Annotation SOP: Defined a "port-sleeve" mask that excludes the central lumen, ensuring geometric consistency for downstream computer vision tasks.
2. Large-Scale Dataset: Created Cholec80-port, a large-scale dataset with 38k+ frames and 1,398 positive port samples, derived from Cholec80 and unified with cleansed versions of m2caiSeg and GynSurg.
3. Baseline Model & Tooling: Released a pretrained segmentation model and associated tooling to support future research in surgical scene understanding.

5. Significance and Future Work

• Significance: This work addresses a critical gap in surgical computer vision by providing a dataset specifically optimized for geometry-based pipelines. By preventing the masking of valid anatomical features (the lumen), it enables more accurate 3D reconstruction and SLAM.
• Limitations: Domain shifts (e.g., transparent ports, varying lighting) still limit universal applicability.
• Future Directions: The authors plan to expand port segmentation coverage across diverse surgical environments and integrate port masking directly into geometry-based pipelines (vSLAM, 3D reconstruction, panoramic inference) to validate the practical utility of the geometric consistency approach.