LAparoscopic Skill and Kinematics is a laparoscopic peg-transfer dataset that pairs endoscopic video with electromagnetic measurements from two instruments, manual instrument annotations, and participant experience metadata.
Release status
Zenodo version 1.0 contains 37 recordings. The complete 115-recording collection used by the quantitative analysis will be added to the same concept DOI. This repository already provides the analysis code, aggregate results, feature definitions, and representative figures. It does not expose participant-level data from recordings that have not yet been released.
- Dataset guide
- Quantitative analysis
- Motion feature definitions
- Phase annotation protocol
- Annotation caveats and recording usage notes
- Analysis script
- Phase-motion analysis script
- Aggregate result tables
- Representative figures
The three cohorts used different collection settings. Raw measurements should therefore be interpreted within cohort or with cohort included explicitly in the statistical model.
| Cohort | Collection | Instrument channels | Frame rate | Public v1.0 | Complete collection |
|---|---|---|---|---|---|
| Paediatric | British Association of Paediatric Endoscopic Surgeons meeting, November 2024 | Position, orientation, relative jaw opening | 13 frames/s | 10 | 30 |
| Urology 1 | Urology boot camp, October 2023 | Position and orientation | 26 frames/s | 8 | 24 |
| Urology 2 | Urology boot camp, October 2024 | Position, orientation, relative jaw opening | 13 frames/s | 19 | 61 |
| Total | 37 | 115 |
The public release is organised as Urology 2 training data, Paediatric
validation data, and Urology 1 testing data. The names 7DOF2024,
BAPES2024, and 6DOF2023 are retained in files for compatibility.
In the analysis and figures, the acquisition channels are described as the
left tool and right tool, according to their usual side of entry in the
endoscopic image. These are image-side labels, not dominant-hand and
non-dominant-hand labels. The mapping was checked against annotated reference
frames from each cohort. Internal fields retain tool1_* and tool2_* names
for compatibility.
- Endoscopic video of a complete peg-transfer attempt.
- Three-dimensional tool position in millimetres for both instruments.
- Tool orientation as a unit quaternion for both instruments.
- A relative jaw-opening signal in the two seven-channel cohorts. This is a per-recording voltage-derived opening fraction, not an absolute jaw angle.
- Manual masks and instrument landmarks on selected video frames.
- Self-reported handedness and laparoscopic procedure experience where collected.
- Dense action-phase labels for the currently annotated subset.
The electromagnetic sensor is mounted near the instrument base and calibrated to estimate tool-tip position. The complete data structure and coordinate conventions are described in the dataset guide.
The companion analysis keeps four denominators separate.
| Analysis unit | Recordings | Study identifiers | Purpose |
|---|---|---|---|
| Complete inventory | 115 | 111 | Describe every available recording |
| Primary motion set | 107 | 107 | Motion inference without repeated identifiers |
| Phase inventory | 38 | 37 | Describe every densely annotated recording |
| Primary phase set | 34 | 34 | Phase comparisons without repeated identifiers |
The 38 phase-labelled recordings contain 425 placement-complete transfer cycles. The primary phase set contains 383 cycles.
The main findings are:
- Four measurements showed novice, intermediate, and expert medians in the experience-associated order in every cohort and retained corrected associations with continuous lifetime procedure volume. They were task duration, left- and right-tool normalised jerk, and right-tool stop-start peaks. Novice-to-expert median reductions were 13--23%, 26--43%, 3--42%, and 13--16%, respectively.
- Eight features retained within-cohort associations with lifetime procedure volume after correction. These included the four ordered measurements, left-tool path length and stop-start peaks, right-tool speed, and bimanual correlation. The bimanual association varied by cohort and is interpreted as a contextual measurement rather than a universal threshold.
- Phase-specific reach and transport measurements separated novice- and expert-labelled recordings with a mean held-out-cohort balanced accuracy of 0.89 after calibration to the unlabelled local cohort and pooled balanced accuracy of 0.83, compared with 0.54 for duration alone. This is an exploratory result from 25 eligible phase-labelled recordings and requires prospective validation.
- Cohort accounted for 88% and 92% of the variation in left- and right-tool speed, respectively. This shows why unadjusted pooling across collection settings is misleading.
- Data-derived lower, middle, and upper motion-score bands had negligible agreement with procedure-count groups (adjusted Rand index 0.019). These are descriptive motion strata, not clinical skill grades.
- Removing duration-dependent normalised jerk retained a related continuous score but changed many band assignments (adjusted Rand index 0.351). None of the six contextual duration or task-efficiency associations then survived correction.
- Models recovered the motion-score rule with high accuracy because the target was calculated from the same motion domains. This is a software consistency check and must not be interpreted as independent skill prediction.
Task duration, movement smoothness, and stop-start peaks are prioritised for prospective feedback validation because their ordered patterns were consistent across all three cohorts within this dataset. Travel distance and coordination between the tools provide useful supporting context. Relative jaw voltage localises instrument actuation during grasp and transfer and is retained as a complementary handling signal. Together, these measurements can describe what changed during a trial rather than assigning performance from procedure count alone.
Full methods, confidence intervals, corrected probability values, sensitivity analyses, and limitations are provided in the analysis guide.
All 107 independent recordings are shown after conversion to relative within-cohort scores. Higher values denote shorter, smoother, or less stop-start performance. The novice, intermediate, and expert medians progress in the experience-associated direction in every panel, while individual points retain the clinically important variation within each experience group.
The same experience measure can have different motion relationships in each
cohort. The top row shows bimanual correlation across all 107 primary
recordings. The lower row uses the phase-labelled subset to show mean cycle
duration. Corrected all-recording results are available in
all_trial_feature_statistics.csv.
Each colour shows the expert-minus-novice effect estimated within one cohort. The wide and sometimes inconsistent intervals show why a pooled effect can hide acquisition-specific uncertainty. The analysis therefore avoids treating the three cohorts as if their coordinate systems and equipment were interchangeable.
The coordination and control score is interpreted continuously. Duration and task efficiency are shown as contextual measurements rather than external clinical validation. A sensitivity analysis without normalised jerk shows why the categorical bands should not be treated as fixed performance grades.
Additional examples include the cohort inventory and phase timing analysis.
Create an environment and install the recorded package versions:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txtThe script uses the sibling BTPN-MT and AI-ELT paths by default. Other
locations can be supplied explicitly:
export LASK_CODE_ROOT=/path/containing/AI-ELT-and-BTPN-MT
export LASK_PHASE_CACHE=/path/to/phase_cache
export LASK_AI_ELT_ROOT=/path/to/AI-ELT
export LASK_ORIGIN_MOTION=/path/to/per_recording_kinematic_json
python scripts/build_scirep_analysis.pyExpected inputs and their schemas are listed in
data/README.md. The current Zenodo release does not yet
contain every cache needed to regenerate the 115-recording paper analysis.
Until the staged release is expanded, the repository provides the aggregate
outputs and file hashes produced by the complete internal collection.
Procedure volume is an experience measure, not a direct assessment of competence. The motion-score bands are relative partitions of two calculated motion domains. They are not scores from the Objective Structured Assessment of Technical Skills, Global Operative Assessment of Laparoscopic Skills, Global Evaluative Assessment of Robotic Skills, McGill Inanimate System for Training and Evaluation of Laparoscopic Skills, or Fundamentals of Laparoscopic Surgery.
One trained researcher annotated the action phases. Participant experience metadata were not visible during annotation. Differences in apparent confidence and fluency could still be perceived from the videos, but informal impressions of skill were not recorded and were not used to define phase labels. Independent phase annotation and expert rating remain planned validation work.
Please cite the dataset using the concept DOI so that the citation resolves to the latest release:
@dataset{Choudhry2026LASK,
title = {LASK: A Dataset for Laparoscopic Skill and 7-DoF Kinematics},
author = {Choudhry, Omar and Jones, Dominic},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.20752650},
url = {https://doi.org/10.5281/zenodo.20752650}
}The original dataset paper and related work are listed in
CITATION.cff.
The dataset and repository contents are released under the Creative Commons Attribution 4.0 licence. For questions, open a GitHub issue or contact Omar Choudhry, School of Computing, University of Leeds.