MAVIS: Robotic Microvascular Anastomosis

• MAVIS is a micro-surgical approach that uses autonomous robotic systems to achieve sub-millimeter precision in connecting small-diameter vessels.
• It employs advanced imaging (OCT, high-speed microcamera) and vision-based algorithms for real-time tissue detection and accurate stitch verification.
• The framework includes comprehensive datasets and hierarchical workflow annotations to benchmark AI models for segmentation, phase recognition, and procedural analysis.

Micro-surgical Artificial Vascular anastomosIS (MAVIS) encompasses research, datasets, and autonomous robotic systems dedicated to the precise artificial connection of small-diameter blood vessels using micro-surgical techniques. This field integrates advances in robotics, computer vision, biomechanical modeling, and workflow understanding to address the inherent technical demands and outcome variability of manual microvascular anastomosis. MAVIS serves as both an application domain—enabling autonomous micro-surgical suturing, coupling, and verification—and a data resource fostering multimodal learning and benchmarking in surgical scene understanding.

1. Definition and Scope

Micro-surgical Artificial Vascular anastomosis ("MAVIS") refers to the surgical joining of microvessels with diameters typically in the 1–3 mm range, performed under microscopic visualization and requiring sub-millimeter accuracy in suture placement and tissue manipulation. Manual microanastomosis is fundamental to procedures such as organ transplantation and reconstructive microsurgery but is constrained by the need for advanced dexterity and extensive training, resulting in revision rates up to 7.9% due to technical error or poor suture quality (Haworth et al., 2024). The MAVIS paradigm extends this task to artificial systems—robotic platforms augmented by imaging, sensing, and algorithmic control—ultimately aimed at reducing operator dependence and standardizing outcome.

Additionally, "MAVIS" designates a comprehensive dataset embedded within the SurgMLLMBench benchmark, supporting pixel-level segmentation, hierarchical workflow annotation, and visual question answering (VQA) for micro-surgical scene understanding (Choi et al., 26 Nov 2025). This dataset focuses on the entirety of the microvascular anastomosis workflow on artificial vessels and serves as a reference for the development and evaluation of surgical AI models.

2. System Architectures and Automation

The state-of-the-art MAVIS implementation is embodied in the Micro Smart Tissue Autonomous Robot (micro-STAR), reported as the first robotic system to autonomously perform vascular anastomosis on real tissue with minimal human intervention (Haworth et al., 2024). Key architectural features are:

• Robotic Manipulator: Seven-DoF KUKA LBR Med arm (sub-mm accuracy, ROS/MoveIt! motion planning).
• Suturing End Effector: Modified Endo360 system, housing a curved-needle driver actuated by dual DC motors; equipped with a 1060 nm OCT fiber-optic probe and a high-speed microcamera (OV6946, 60 fps).
• Vessel Positioning System (MAPS): Bilateral nitinol self-expanding clamps (2–8 mm diameter), dual rotary stages (±180°), and antislip coatings, allowing precise, programmable rotation and grasping of the vessel segment.

The micro-STAR system performs autonomous interrupted suturing through real-time tissue detection (OCT), closed-loop trajectory correction, and vision-based verification of stitch success. Autonomous autonomy rate reached 90% (fraction of sutures placed without human intervention).

The following table summarizes core architectural components:

Component	Function	Specifications
Robotic Arm	Macromanipulation & Positioning	7-DoF KUKA LBR Med, 70 cm workspace
Suturing Tool (Endo360)	Needle Driving, Imaging	30 mm radius, integrated OCT + microcamera
MAPS Vessel Holders	Accurate Grasp & Rotation	Nitinol holders, ±180° rotation, 2–8 mm range
Sensing	Intraoperative Feedback	OCT: ≤50 µm axial resolution, 100 kHz A-scans
Vision	Suture Verification	ResNet-50, 6-channel, F1 ≈ 0.88

Key control elements include real-time ROS/ROS2 integration, MoveIt! trajectory planning with needle approach constraints, and feedback fusion which prioritizes OCT for depth correction but reserves vision feedback exclusively for detecting missed sutures (Haworth et al., 2024).

3. Sensing, Vision, and Perception Algorithms

Robust autonomous MAVIS requires a multi-modal sensing approach:

• OCT-Based Tissue Boundary Estimation: One-dimensional A-scan acquisition at 100 kHz identifies tissue/nitinol/air transitions via template matching and normalized RMSE scoring. The leading matched window determines $z_\text{edge}$, governing bite-depth offsets for suture placement.
• Vision-Based Suture Verification: Modified ResNet-50 operates on concatenated "before" and "after" RGB images to classify missed sutures, with retrial prompted at $P_\text{missed} > 0.5$. Tested accuracy is 87.0%, F1-score ≈ 0.88.

Advanced suture and thread detection, crucial for monitoring multi-step workflows, leverages bottom-up deep learning frameworks. The multi-stage system proposed by Jin et al. uses cascaded FCNs to generate a gradient map $G(x,y)$ and overlap map $O(x,y)$ for 2D suture localization, achieving PSNR ≈ 31.9 dB and Overall Thread Tracking Precision (OTTP) ≈ 1.33 px for marked sutures (Hu et al., 2017). For MAVIS, these architectures require higher input resolutions, lightweight attention backbones, and topology-aware loss functions to adapt to subpixel wire localization and real-time constraints (≥30 fps) typical of micro-surgical environments.

The following summarizes the multi-stage deep suture detection pipeline:

Step	Method	Result
1. Map	Stage 1 FCN (ResNet/UNet) for $G(x,y)$, $O(x,y)$	Initial thread centerline + overlap
2. Refine	Stage 2 FCN with input fusion	Smoothed, refined gradient map
3. Detect	Curvilinear structure detector, polarity and region growing	Ordered centerline points
4. Fit	Segment linking, monotonicity analysis, cubic spline fitting	Final thread centerline reconstruction

Adaptations for MAVIS include model pruning, mixed-precision inference, and swap-in of learned centerline detectors for conventional line filters, as well as temporal smoothing (ConvLSTM, Kalman filter) for motion robustness in occluded workspace (Hu et al., 2017).

4. Datasets and Hierarchical Workflow Annotation

The MAVIS dataset, part of SurgMLLMBench, constitutes the most comprehensive micro-surgical vascular anastomosis video dataset to date (Choi et al., 26 Nov 2025). It provides:

• Data Acquisition: 19 videos across three expert surgeons, each capturing the complete anastomosis workflow on 1 mm artificial vessels at 1920×1080 resolution, annotated at 1 fps.
• Hierarchical Labels: Three-level taxonomy per frame (Stage, Phase, Step), mapping complete multi-step procedures.
• Pixel-Level Segmentation: 8-class manual polygon masks (background, forceps, scissors, vascular clamps, needle holder, vessel, needle, thread), with instance separation for overlapping tools.
• VQA Annotation: Five template question-answer pairs per frame, covering workflow query, tool count, instrument type, instrument action, and dataset source; tightly linked to instance segmentation.
• Statistics: 10 652 frames, with Stage, Phase, Step class distributions recorded to expose overrepresented (suture/knot-tying) and rare (cutting, flip) events. Vessel and thread are the most frequent segmented classes.

No fixed train/val/test split is provided; cross-surveillance splits (by surgeon or video) are recommended to avoid subject leakage. The dataset is intended for benchmarking multimodal AI models for segmentation and workflow recognition.

5. Performance Evaluation and Benchmarking

Robotic MAVIS system performance is evaluated through both experimental and algorithmic metrics:

• Physical Anastomosis Quality: Leak pressure, lumen reduction, and suture placement variation (COV%) assessed on 5 mm porcine femoral arteries. μSTAR demonstrated leak pressure (0.32±0.23 PSI) statistically indistinguishable from human surgeons; bite-depth and spacing COV% matched or exceeded surgeon consistency, with an autonomy rate of 90%. However, μSTAR was 2–4× slower per stitch (353±40 s) compared to surgeons (p<0.01) (Haworth et al., 2024).
• Dataset Benchmarks: On SurgMLLMBench, LLaVA instruction-tuned on other surgical domains transferred with stage recognition accuracy ≈ 75% after fine-tuning on MAVIS, while segmentation models exhibited domain-specific drops for microscope imagery (Choi et al., 26 Nov 2025).
• MISAW Comparison: Previous micro-surgical video challenges, e.g., MISAW, achieved >95% AD-Accuracy for phase recognition and ≈80% for step-level; activity recognition (atomic, instrument-specific events) was limited to ≈61%, underscoring the need for richer annotation and algorithmic development as now being addressed in MAVIS (Huaulmé et al., 2021).

Key dataset segmentation and recognition metrics:

Metric	Formula	Context
IoU	$TP_c / (TP_c + FP_c + FN_c)$	Per-class Intersection over Union
mIoU	$(1/C) \sum_{c} \text{IoU}_c$	Mean IoU across C classes
Dice	$2 TP_c / (2 TP_c + FP_c + FN_c)$	Dice coefficient (not explicit)
AD-Acc	$(1/C)\sum_{c=1}^{C} TP_c/(TP_c+FN_c+FP_c)$	Balanced activity accuracy (MISAW)

6. Biomechanical Modeling and Coupler Design

MAVIS also encompasses engineering and fluid dynamic modeling for artificial couplers used in microvascular anastomosis. Analytical solutions to pulsatile blood flow through coupled microvessels reveal that:

• Wall Protrusion/Intrusion: Axisymmetric coupling devices (protrusions) reduce peak vessel wall shear strain by up to ≃15%, whereas sutured intrusions increase it by ≃15%.
• Gradient of Wall Deformation: The shear gradient is modulated by the geometric curvature parameter $\beta$: larger $\beta$ yields a flatter gradient, lowering the local thrombosis risk. A geometric design with bump amplitude $\epsilon\approx0.03$ and $\beta \gtrsim 0.3$ keeps peak shear-rate gradients below 10% of unperturbed values.
• Trade-Offs: Increased $\beta$ (gentler curvature) improves hemocompatibility at the cost of mechanical grip. Excessive wall compliance yields unpredictable deformations and potential shear spikes (Gallagher, 2018).

Table: Mechanical design recommendations for MAVIS couplers (from (Gallagher, 2018)):

Parameter	Recommendation	Design Rationale
Bump amplitude	$\epsilon$ ≈ 0.03–0.05	Limit wall shear perturbation
Curvature $\beta$	$\beta \gtrsim 0.3$	Flatten gradient of wall deformation
Wall compliance	Small, controlled	Prevent unpredictable shear spikes

7. Challenges, Limitations, and Future Directions

MAVIS platforms face several outstanding obstacles:

• Miniaturization: Current mechanisms (e.g., Endo360) are sized for 3-0 suture; MAVIS requires handling of 8-0–10-0 monofilament, necessitating further miniaturization of drive and optical subsystems (Haworth et al., 2024).
• Automation Gaps: Suture pull-through/cutting and knot-tying remain manual; full autonomy would require automated knotting or coupler placement.
• Mechanical Robustness: Vessel slippage on nitinol holders and tangential load response require improved clamp engineering, possibly force-feedback integration.
• Workflow Complexity: Activity recognition accuracy remains suboptimal (<60%) for highly fine-grained actions, necessitating new models, temporally-aware architectures, and domain-adapted pretraining (Huaulmé et al., 2021).
• Regulatory Translation: Sterilization protocols and disposable imaging sleeves will be required for clinical deployment; animal models and first-in-human trials are projected as translational steps.

Datasets such as MAVIS will enable the development and benchmarking of AI and robotic systems in high-fidelity, multimodal micro-surgical scenarios, supporting progress toward fully autonomous, outcome-optimized microvascular anastomosis platforms suitable for broad clinical adoption.