GynSurg: Surgical Data & AI Ecosystem

Updated 3 July 2026

GynSurg is an extensive ecosystem featuring annotated high-definition surgical videos, expert consensus, and machine learning benchmarks for gynecologic laparoscopic procedures.
It enables real-time decision support and post-operative analysis using CNN-RNN hybrids and transformer networks to achieve high accuracy and robust phase segmentation.
The platform employs rigorous evaluation metrics like accuracy, F1-score, and mIoU while addressing challenges such as annotation cost, domain generalization, and real-time performance.

GynSurg denotes a domain and an ecosystem of datasets, computational methods, and intelligent systems designed to analyze, understand, and support gynecologic laparoscopic surgery through surgical video data, annotation resources, machine learning benchmarks, and integrated clinical workflows. The term encompasses comprehensive datasets—most notably the GynSurg dataset—as well as associated tools and methodologies that enable surgical workflow analysis, real-time guidance, post-operative assessment, and computer-assisted interventions in gynecologic minimally invasive surgery.

1. Dataset Foundations and Scope

GynSurg is anchored by the publicly released “GynSurg” dataset, the largest and most diverse resource dedicated to gynecologic laparoscopic surgery (Nasirihaghighi et al., 12 Jun 2025, Nasirihaghighi, 13 Aug 2025). This curated collection includes 152 high-definition surgical videos (1920×1080 px, 30 fps) with granular temporal and spatial annotations and an additional set of 75 unlabeled videos for self-supervised and semi-supervised learning workflows. The annotated corpus supports a variety of procedures (laparoscopic hysterectomy, myomectomy, oophorectomy, adnexectomy, lymphadenectomy) and spans >80 hours of operating room footage.

Annotation focuses on:

Temporal segmentation: Delineation of surgical phases (e.g., trocar insertion, dissection, extraction) and fine-grained actions (needle passing, coagulation, suction/irrigation, transection, rest).
Pixel-level masks: Semantic segmentation for 21 instrument types (consolidated to 7 final classes), 4 anatomic structures (uterus, fallopian tube, ovary, auxiliary organs), plus granular organ and tool presence (Nasirihaghighi et al., 12 Jun 2025).
Event markers: Binary labels for intraoperative side effects (bleeding, smoke) and occurrence of clinically critical events (e.g., abdominal access, needle passing) as described in related datasets (Nasirihaghighi et al., 2023).

Three board-certified gynecologic surgeons performed iterative expert annotation, reaching consensus with Cohen’s κ ≈ 0.85 for phases and κ ≈ 0.80 for actions, ensuring high reliability (Nasirihaghighi, 13 Aug 2025).

2. Benchmark Tasks, Taxonomies, and Evaluation Protocols

The GynSurg ecosystem supports multi-task evaluation:

Action recognition: Five-class clip-level classification of major surgical actions (coagulation, needle-passing, suction/irrigation, transection, rest), using temporal clips (3 seconds, 1 s stride) (Nasirihaghighi et al., 12 Jun 2025).
Phase segmentation: Frame-level classification of surgical stages.
Event detection: Temporal boundary localization for events such as bleeding and needle passing, with rigorous clip-level annotation and overlap (Nasirihaghighi et al., 2023).
Semantic segmentation: Multi-class pixel-wise instrument and anatomy identification, supporting both primary and auxiliary tool categories.
Side-effect binary detection: Bleeding and smoke presence detection per frame.

Standard metrics are enforced, including accuracy (Acc), F1-score, mean intersection-over-union (mIoU) for segmentation, and mean average precision (mAP) for detection tasks:

$\mathrm{Acc} = \frac{TP + TN}{TP + TN + FP + FN}$

$\mathrm{F1} = \frac{2TP}{2TP + FP + FN}$

$\mathrm{mIoU} = \frac{1}{K} \sum_{k=1}^K \frac{TP_k}{TP_k + FP_k + FN_k}$

$\mathrm{mAP} = \int_{0}^{1} P(r)\,dr$

where $TP$ , $TN$ , $FP$ , $FN$ denote counts of true/false positives/negatives, $K$ is the number of segmentation classes, and $P(r)$ is precision at recall $\mathrm{F1} = \frac{2TP}{2TP + FP + FN}$ 0 (Nasirihaghighi et al., 12 Jun 2025, Nasirihaghighi, 13 Aug 2025).

Cross-validation is performed at the video level, with four- or five-fold splits, carefully avoiding segment leakage between training and test sets (Nasirihaghighi et al., 12 Jun 2025).

3. Methods and Learning Paradigms

GynSurg provides benchmarks for a broad spectrum of video understanding and image analysis models:

CNN–RNN hybrids: Stacked BiLSTM/GRUs following strong CNN backbones (e.g., ResNet50), processing temporal frame features for clip-wise classification. These architectures deliver up to 86.8% mean accuracy and >90% F1 for main actions (Nasirihaghighi et al., 2023).
Transformer networks: TimeSformer and hybrid models, used especially for surgical phase and event recognition, consistently outperforming RNN-based approaches by leveraging inter-frame temporal dependencies via multi-head self-attention, with state-of-the-art results on event detection tasks (average accuracy 86.10%, F1 86.03%) (Nasirihaghighi, 13 Aug 2025, Nasirihaghighi et al., 2023).
Semantic segmentation frameworks: DeepLabV3 (VGG16, ResNet), UPerNet, and SAM-PP for instrument and organ masking, achieving Dice coefficients up to 83.35% for forceps, 85.73% for sealer-dividers, and 82.89% mean for SAM-PP (Nasirihaghighi et al., 12 Jun 2025).

Recent research emphasizes data-efficient and semi-supervised learning, introducing methods such as DIST, SemiVT-Surg, and ENCORE to exploit unlabeled video and minimize annotation requirements. These leverage dual invariance self-training, contrastive prototype loss, temporal consistency regularization, and adaptive class-aware thresholding. Gains of +4–10 percentage points in low-label regimes, and improved cross-site/procedure robustness, are consistently reported (Nasirihaghighi, 13 Aug 2025).

4. Datasets and Resources Supporting GynSurg

Multiple datasets underpin GynSurg development:

Dataset	Scope	Annotations	Access/License
GynSurg	152+ videos; 12,362 frames	Phase/action/event/action/segmentation	CC BY-NC-ND 4.0 (Nasirihaghighi et al., 12 Jun 2025)
GLENDA	25,682 frames; 400+ videos/segments	Endometriosis lesion masks/class labels	CC-BY-NC (Leibetseder et al., 29 Aug 2025)
LapGyn6-Actions	18 videos	Action-class expert annotation; 7-class binary	Code online (Nasirihaghighi et al., 2023)
Event Recognition (Nasirihaghighi et al., 2023)	174 videos	4 surgical events annotated, clinical review	See paper

GLENDA focuses on endometriosis lesion detection, with pixel-precision masks and multi-class lesion taxonomy, used extensively for benchmarking segmentation models (e.g., Mask R-CNN, U-Net variants). Event and action recognition datasets supplement GynSurg’s scope with additional event-centric temporal annotations (Leibetseder et al., 29 Aug 2025, Nasirihaghighi et al., 2023).

5. Applications and Clinical Integration

GynSurg enables the development, benchmarking, and deployment of systems for:

Intraoperative decision support: Real-time detection of bleeding, smoke, instrument type, and critical events, facilitating alerting and workflow guidance (Nasirihaghighi et al., 12 Jun 2025, Nasirihaghighi et al., 2023).
Postoperative analysis: Automated video documentation, key-frame extraction, residual lesion quantification, and derivation of objective skill metrics (Leibetseder et al., 14 Oct 2025).
Training and education: Creation of annotated video libraries for typical and rare procedural events, interactive quizzes, and data-supported skill assessment (Leibetseder et al., 29 Aug 2025).
Research in surgical workflow understanding: Enables multi-task and self-/semi-supervised learning, domain-adaptive methods, and discovery of novel procedural patterns (Nasirihaghighi, 13 Aug 2025).

Several demo systems illustrate specific integration pathways, such as real-time Mask R-CNN overlays for endometriosis detection (5–6 fps), post-surgical video annotation with export for archives, and automated surgical action logging (Leibetseder et al., 14 Oct 2025, Nasirihaghighi et al., 12 Jun 2025).

6. Extension to Robotic Surgery and Scene Modeling

GynSurg resources interconnect with advanced scene reconstruction methods (e.g., SurgicalGaussian (Xie et al., 2024)), supporting real-time deformable 3D modeling of surgical fields for applications in robotic guidance and intraoperative navigation. These approaches use depth/mask-guided 3D Gaussian splatting, regularization for tissue and tool geometry, deformation-aware rendering, and mask-driven supervision to achieve high fidelity and real-time rates (e.g., 80–140 fps, 0.89–0.97 SSIM in benchmarks). Such methods augment GynSurg’s utility in robot-assisted minimally invasive surgery and simulation (Xie et al., 2024).

Fully automated systems (e.g., Nahid framework (Saadati, 2023)) demonstrate closed-loop diagnosis and treatment, fusing segmentation, path planning, and robotic actuation, further extending GynSurg’s impact beyond passive video understanding to interventional automation and surgical robotics.

7. Limitations, Challenges, and Future Directions

Despite its comprehensiveness, GynSurg exhibits certain limitations:

Dataset bias and coverage: While large and diverse, rare anatomic variants, infrequent complications, and atypical procedures remain under-represented (Nasirihaghighi et al., 12 Jun 2025).
Annotation cost and scalability: Expert annotation is expensive and time-consuming. Semi-supervised and self-supervised methods partially mitigate but do not eliminate this constraint (Nasirihaghighi, 13 Aug 2025).
Generalization: Models trained on GynSurg demonstrate 8–12 percentage points F1-score drop under cross-hospital/procedure evaluation; advanced semi-supervised and domain-adaptive frameworks reduce this but do not eliminate the domain gap (Nasirihaghighi, 13 Aug 2025).
Real-time constraints: Segmentation and detection models occasionally lag clinical video rates (e.g., 6 fps for Mask R-CNN segmentation on standard GPUs), motivating quantization, pruning, and model compression research (Leibetseder et al., 14 Oct 2025).

Future research directions include richer multi-class and multi-structure segmentation, transformer-based temporal modeling, temporal consistency mechanisms, interactive correction GUIs, and more robust domain-adaptive pipelines. Further clinical studies are required for prospective validation and regulatory approval in live intraoperative contexts (Leibetseder et al., 14 Oct 2025).

GynSurg, supported by datasets, learning paradigms, and integrated clinical feedback, constitutes the foundational infrastructure for rigorous, data-driven development and deployment of intelligent systems in gynecologic laparoscopic surgery, and provides critical benchmarks and modeling tools for the broader surgical computing community.