Camera-Based Metrics Show Surgeons Really Do Suture Better, Study Finds

Every surgeon knows the moment: the needle driver dips into a deep cavity, the thread follows, and the quality of that single stitch may determine whether a vascular repair holds or fails. For decades, judging whether a trainee can perform that maneuver reliably has depended on an experienced educator standing over the bench, watch in hand, filling out a rating form. It is subjective, slow, and increasingly difficult to scale at a time when the rise of endovascular techniques means surgical residents encounter open procedures far less often than their predecessors did. A new study published in Global Surgical Education, the journal of the Association for Surgical Education, offers a strikingly different vision: a simulator that watches every stitch through a camera and reduces suturing skill to a set of numbers, no human rater required.

The research, led by Jianxin Gao and Richard E. Groff of Clemson University’s Department of Electrical and Computer Engineering together with colleagues in bioengineering, vascular surgery, and medicine, reports the development and large-scale evaluation of image-based metrics for assessing open surgical suturing skill. The team built a simulator called SutureCoach, designed around the clock-face radial suturing task used in the Fundamentals of Vascular Surgery examination, a simulation-based assessment developed to address the shrinking exposure of trainees to open vascular procedures. In that task, participants place sutures around a circular membrane in a counter-clockwise pattern, mimicking the geometry of a vascular anastomosis. What makes SutureCoach distinctive is not the task itself but the instrumentation: a camera mounted inside the membrane housing records the needle and thread from below as they pass through the material being sutured.

That under-membrane camera is the heart of the system. Recording RGB video at 640 by 480 resolution and 60 frames per second, illuminated by an LED strip behind a diffusing panel to guarantee stable lighting, it feeds a computer vision program that tracks the needle tip and body in real time as the needle pierces the membrane. Because the camera sits beneath the suture plane, the system sees precisely what matters most: the trajectory of the needle through the tissue surrogate, the angle at which it enters, the path it follows, and the way it is withdrawn. Crucially, no sensors need to be attached to the surgeon’s hands or instruments. The participant simply sutures, and the software does the rest. A second camera above the workspace captures RGB and depth video, while electromagnetic trackers, inertial measurement units, and a force sensor round out the simulator’s sensor suite, though the new study focuses specifically on what the cameras alone can reveal.

The technical core of the paper lies in how the software dissects each suture. A refined phase-segmentation scheme divides every stitch into four phases: approach, insertion, removal, and pull-out. The approach phase covers everything before the needle punctures the membrane. The insertion phase spans the moment the needle body enters the housing until the tip moves toward its exit point. The removal phase begins when the participant re-grasps the needle tip and ends when the needle body clears the housing, and the pull-out phase covers drawing the thread through. The software identifies three timestamps to make these divisions automatically: tip entry, when the needle is first detected; tip exit, when the needle tip begins piercing upward through the membrane; and swage exit, the last frame in which the needle is visible. This phase structure matters because different phases probe different aspects of skill, and because it allows metrics to be computed over precisely defined windows of the task.

From this segmentation, the researchers evaluated eight image-based metrics. Four were carried over from earlier work on a simulator prototype: needle tip path length, which measures motion economy during insertion; needle swept area, the accumulated area covered by the needle as it moves through insertion and removal; needle sway length, which quantifies how much the needle rotates about the axis from entry point to tip; and stitch length, the simple distance between entry and exit points. Four new metrics were inspired by the refined phase segmentation. Three of them, the range and average of the needle angle and the range of the orthogonal tip projection, capture lateral needle movement during insertion, on the principle that skilled suturing involves inserting the needle perpendicular to the tissue surface and following its natural curvature. The fourth, removal time, measures how efficiently the participant extracts the needle. Each metric translates a visually observable behavior into a number that can be compared across individuals.

To test whether these numbers actually mean anything, the team ran a study of unusual scale for this field. Over five months, at a university campus, a hospital, and several vascular conferences, 97 participants performed four trials each on SutureCoach, alternating between a surface condition and a depth condition in which clear cylindrical constraints forced the suturing to occur inside a cavity. The participants divided almost evenly into three groups: 35 experts, meaning fellows and attending surgeons; 32 resident surgeons spanning postgraduate years one through five; and 30 novices, mostly medical students. The resulting dataset contained more than 4,700 detected individual sutures, with fewer than half a percent discarded due to failed needle detection. All trials used 3-0 Prolene suture with a 26-millimeter half-circle needle, and all participants received standardized instructions before starting.

The statistical results were unambiguous at the level that matters most for basic competence. One-way analysis of variance showed that all eight metrics had significantly different group means when comparing novices with surgeons, at both the surface and depth conditions. In other words, a camera watching the needle from below can reliably tell a medical student from a trained surgeon without ever seeing the surgeon’s hands. Welch’s t-tests confirmed the pattern in pairwise comparisons, and receiver operating characteristic analysis showed that the metrics carry genuine classification power, not just statistical significance. The needle swept area metric proved the strongest single classifier for distinguishing novices from surgeons at depth, achieving the largest area under the ROC curve among the eight metrics.

The more demanding question was whether the metrics could separate residents from attending surgeons, two groups whose performances differ far more subtly. Here the depth condition proved decisive. At the surface, most metrics could not statistically distinguish the two surgeon groups. But when suturing inside the cylindrical cavity, seven of the eight metrics showed significant mean differences between residents and attendings. The researchers attribute this to the added complexity of suturing at depth, which amplifies the differences in technique that experience produces. The boxplot data reinforced the point: attending surgeons showed the smallest spread of metric values at depth for six of the eight metrics, indicating more consistent performance, and recorded the smallest median and mean values for swept area, sway length, needle angle range, average needle angle, and orthogonal tip projection range, all signs of reduced lateral needle movement. They were also faster in the removal phase.

The study is candid about its limits. Clinical standing served as a proxy for true suturing skill, which is imperfect, particularly near the fuzzy boundary between senior residents and junior attendings where years of experience cluster. Specular reflections from the needle occasionally corrupted tip detection, which the authors believe blunted the performance of the tip path length metric compared with an earlier pilot study, and motivated the light diffuser added to the current simulator. Recruitment at vascular conferences may have introduced selection bias toward residents already interested in the field. Still, the implications are considerable. An objective, camera-based assessment that requires no instrumented tools and no human rater could be deployed wherever a simulator fits, providing standardized feedback at a moment when open surgical experience is becoming scarce. The work was supported by the National Institutes of Health under grant 5R01HL146843, and the authors suggest that future versions will incorporate independent expert ratings as ground truth and alternative membrane materials to eliminate the residual detection failures. If those steps succeed, the humble stitch may soon be graded not by a watchful professor but by an unblinking lens.

Subject of Research: Image-based computer vision metrics for objective assessment of open surgical suturing skill

Article Title: Developing and evaluating image-based metrics for open surgical suturing skill assessment

Article References: Developing and evaluating image-based metrics for open surgical suturing skill assessment. (n.d.). https://doi.org/10.1007/s44186-026-00596-2

Image Credits: AI Generated

DOI: 10.1007/s44186-026-00596-2

Keywords: surgical skill assessment, computer vision, suturing simulator, vascular surgery, medical training, SutureCoach, image-based metrics, simulation-based education, needle tracking, objective assessment, open surgery, machine learning