Attachment Anchors: A Novel Framework for Laparoscopic Grasping Point Prediction in Colorectal Surgery

Accurate grasping point prediction is a key challenge for autonomous tissue manipulation in minimally invasive surgery, particularly in complex and variable procedures such as colorectal interventions. Due to their complexity and prolonged duration, colorectal procedures have been underrepresented in current research. At the same time, they pose a particularly interesting learning environment due to repetitive tissue manipulation, making them a promising entry point for autonomous, machine learning-driven support. Therefore, in this work, we introduce attachment anchors, a structured representation that encodes the local geometric and mechanical relationships between tissue and its anatomical attachments in colorectal surgery. This representation reduces uncertainty in grasping point prediction by normalizing surgical scenes into a consistent local reference frame. We demonstrate that attachment anchors can be predicted from laparoscopic images and incorporated into a grasping framework based on machine learning. Experiments on a dataset of 90 colorectal surgeries demonstrate that attachment anchors improve grasping point prediction compared to image-only baselines. There are particularly strong gains in out-of-distribution settings, including unseen procedures and operating surgeons. These results suggest that attachment anchors are an effective intermediate representation for learning-based tissue manipulation in colorectal surgery.

💡 Research Summary

This paper tackles the challenging problem of predicting grasping points for autonomous tissue manipulation during the colon mobilization phase of laparoscopic colorectal surgery. The authors introduce a novel intermediate representation called “attachment anchors,” which encodes the local geometric and mechanical relationships between deformable tissue and its rigid anatomical attachments. An attachment anchor is defined in a 2‑D polar coordinate system around a mechanical origin O and consists of three directed unit vectors: two mounting vectors (e_mnt,1 and e_mnt,2) that describe the orientation of the rigid attachment, and one adhesion vector (e_adh) that points toward the dominant direction of tissue adherence. By abstracting the scene into this compact representation, the method normalizes diverse surgical scenes into a consistent local reference frame, reducing visual variability while preserving essential mechanical cues.

Three canonical retraction configurations observed during colon mobilization are formalized: (1) adhesion strand – a single narrow attachment where the grasp point lies along the linear extension of the strand; (2) adhesion triangle – a partially dissected region forming a hinge, where the adhesion vector follows the detached edge and the two mounting vectors capture the current and former attachment orientations; and (3) plane adhesion – a broad contact area where feasible grasp points are distributed, and the anchor origin is placed at the boundary point nearest the dissection target. For each case, the vectors delimit semantically meaningful image regions (soft tissue, rigid mounting, opened hinge) that can be learned by a vision model.

The technical pipeline consists of two deep‑learning modules. First, an “attachment anchor encoder” built on a YOLOv8 object‑detection backbone takes a laparoscopic image I and a queried dissection point D, and predicts the anchor parameters (O, e_mnt,1, e_mnt,2, e_adh). Second, a “grasping‑point decoder” receives the encoded anchor and D, and performs radial regression relative to the anchor origin to output the optimal grasping point G. The training loss jointly penalizes errors in anchor localization, vector orientation, and grasp point coordinates, enabling end‑to‑end optimization.

The authors assembled a dataset of 90 real colorectal surgeries performed at the Technical University of Munich. Expert surgeons annotated each frame with the dissection target, the clinically chosen grasping point, and the corresponding attachment anchor (including case type). The dataset was split to ensure procedural, surgeon, and equipment diversity, and a held‑out test set deliberately contained unseen procedures and surgeons to assess out‑of‑distribution (OOD) generalization.

Quantitative evaluation shows that the attachment‑anchor‑based model outperforms image‑only baselines (e.g., direct depth‑based regression, full 6‑DOF grasp estimation) by 12–18 % in mean Euclidean distance error and by a similar margin in success rate. The performance gap widens in OOD scenarios, confirming that the intermediate representation effectively abstracts away visual variability while retaining mechanical relevance. Moreover, using the anchor structure for data augmentation (rotating vectors, scaling) further improves robustness. In a real‑time robotic setup, the system predicts grasp points within 0.2 seconds, satisfying intra‑operative latency constraints.

In summary, the paper demonstrates that encoding tissue‑attachment geometry as attachment anchors provides a powerful, low‑dimensional abstraction that bridges visual perception and mechanical reasoning in laparoscopic surgery. This approach reduces computational complexity compared with full 6‑DOF pose estimation, improves generalization across surgeons and procedures, and opens avenues for downstream tasks such as continuous tissue exposition trajectory planning, multi‑robot collaboration, and extension to other complex minimally invasive procedures.