Endoscapes-CVS201: CVS Benchmark in Laparoscopy
- Endoscapes-CVS201 is a benchmark dataset defining the Critical View of Safety with expert multi-label annotations for laparoscopic cholecystectomy.
- It compiles annotations from 201 videos and 11,090 frames, supporting diverse tasks including detection, segmentation, and scene-graph transfer.
- The dataset drives advances in anatomy-aware reasoning and multi-modal CVS prediction while addressing challenges like sparse labels and low inter-annotator agreement.
Endoscapes-CVS201 is the Critical View of Safety (CVS) assessment subset of the broader Endoscapes benchmark for laparoscopic cholecystectomy. It is centered on frame-level, multi-label recognition of the three Strasberg CVS criteria from dissection-phase video frames, and it has become a reference dataset for automated CVS prediction, anatomy-aware reasoning, and mixed-supervision surgical scene understanding. In the official Endoscapes report, the dataset is defined over 201 laparoscopic cholecystectomy videos with 11,090 frames annotated for CVS by three clinical experts, alongside companion detection and segmentation subsets derived from the same cases (Murali et al., 2023). Later work uses Endoscapes-CVS201 as the exclusive downstream dataset for multi-modal CVS recognition and as a downstream benchmark for scene-graph transfer, underscoring its status as a canonical CVS test-bed (Baby et al., 7 Jul 2025, Shin et al., 21 Jul 2025).
1. Dataset identity within the Endoscapes benchmark
Endoscapes-CVS201 denotes the CVS-focused subset of Endoscapes built from 201 laparoscopic cholecystectomy videos, corresponding to 201 procedures and, in the technical report, 201 patients (Murali et al., 2023). Its task is frame-level CVS classification over sparsely but systematically sampled dissection-phase images. The broader Endoscapes release is organized into three official subsets: Endoscapes-CVS201 for CVS classification, Endoscapes-BBox201 for object detection, and Endoscapes-Seg50 for released segmentation; a larger Endoscapes-Seg201 is discussed as a private or ceiling segmentation set in benchmark reporting (Murali et al., 2023).
The nomenclature reflects both task and source cohort. The suffix “201” refers to the 201 laparoscopic cholecystectomy videos from which all benchmark subsets are derived (Murali et al., 2023, Shin et al., 21 Jul 2025). Endoscapes-CVS201 is therefore not an isolated image collection but the CVS-labeled view of a multi-task surgical dataset spanning image-level labels, bounding boxes, and segmentation masks. Earlier work on TCNN already used Endoscapes as a 201-video laparoscopic cholecystectomy dataset focused on the hepatocystic dissection phase relevant to CVS, although that paper did not yet formalize the later “Endoscapes-CVS201” naming convention (Alapatt et al., 2021).
Within this benchmark family, Endoscapes-CVS201 is the backbone for CVS prediction. It is explicitly described as a benchmark for automated Critical View of Safety assessment in laparoscopic cholecystectomy, and subsequent papers treat it as the standard downstream benchmark for both image-only and structure-aware models (Murali et al., 2023, Baby et al., 7 Jul 2025).
2. Clinical target and label formalism
The dataset operationalizes Strasberg’s Critical View of Safety as a binary multilabel problem with three independent but clinically related criteria. The technical report names them as C1 “Two Structures,” C2 “Hepatocystic Triangle Dissection,” and C3 “Cystic Plate” (Murali et al., 2023). A later multi-modal study spells out the frame-level criterion texts as follows: Criterion 1 is “the cystic duct and the cystic artery, connected to the gallbladder,” Criterion 2 is “a hepatocystic triangle cleared from fat and connective tissues,” and Criterion 3 is “the lower part of the gallbladder separated from the liver bed” (Baby et al., 7 Jul 2025).
Formally, each annotated frame carries a three-dimensional binary label vector,
In the benchmark formulation, CVS is thus a multi-label classification problem rather than a mutually exclusive multi-class problem (Murali et al., 2023, Baby et al., 7 Jul 2025). A frame can satisfy multiple criteria simultaneously, and frame-level “overall CVS” is achieved if and only if all three criterion labels are positive. The technical report further defines a video-level aggregation in which a criterion is achieved for a video if it appears in at least one annotated frame, and video-level overall CVS requires that each criterion be achieved at least once somewhere in the video: This distinction between frame-level conjunction and video-level existence is important for interpreting benchmark metrics and split stratification (Murali et al., 2023).
The label design is clinically motivated but statistically difficult. Later work emphasizes that CVS assessment remains challenging even for experts, citing low inter-annotator agreement and using that difficulty to justify ambiguity-aware learning objectives (Baby et al., 7 Jul 2025). This suggests that Endoscapes-CVS201 is not merely a standard multi-label image dataset; it encodes a clinically consequential target whose definition depends on subtle anatomical exposure and dissection state.
3. Sampling regime, annotation protocol, and official splits
All Endoscapes-CVS201 labels are drawn from the dissection phase of the 201 laparoscopic cholecystectomy videos (Murali et al., 2023). The technical report states that the full dissection-phase pool contains 58,813 frames and that CVS labels are sampled sparsely rather than densely: one frame every 5 seconds over the “CVS-evaluable” region, defined as the subset of the dissection phase beginning when any CVS criterion becomes evaluable and ending when the first clip is applied to the cystic duct or artery (Murali et al., 2023). The result is a CVS-labeled set of 11,090 annotated frames (Murali et al., 2023, Baby et al., 7 Jul 2025).
Each CVS-labeled frame is annotated independently by three expert clinicians in the official report, and the benchmark ground truth is defined by majority vote per criterion (Murali et al., 2023). A later multi-modal study describes the labels as confidence scores from three annotators and states that these scores are rounded to 0 or 1 for fair comparison with baselines (Baby et al., 7 Jul 2025). Taken together, these accounts indicate that the dataset preserves multi-expert judgment but that downstream works may instantiate slightly different label post-processing conventions for benchmarking.
The official split is performed at the video level, with no frame overlap across train, validation, and test sets. The technical report gives 120 training videos, 41 validation videos, and 40 test videos, with stratified random sampling based on video-level CVS achievement (Murali et al., 2023). For CVS-labeled frames, the report lists 6,970 training frames, 2,331 validation frames, and 1,799 test frames, whereas a later study using the same official split reports 6,960 training frames, 2,331 validation frames, and 1,799 test frames (Murali et al., 2023, Baby et al., 7 Jul 2025). All sources agree on the total of 11,090 CVS-annotated frames.
The benchmark also defines low-label variants for CVS-only settings. The technical report describes 12.5% and 25% training-video budgets, corresponding to 15 or 30 training videos, with three random subsets per budget (Murali et al., 2023). These reduced-label splits were designed for label-efficient comparison and link Endoscapes-CVS201 to broader questions in mixed supervision and semi-supervised surgical learning.
4. Relationship to detection, segmentation, and scene-graph annotations
A defining characteristic of Endoscapes-CVS201 is that it is embedded in a multi-task annotation ecosystem. Endoscapes-BBox201 uses the same 201 videos and the same dissection-phase frame pool, with 1,933 frames annotated by bounding boxes for six classes: Cystic Plate, HC Triangle Dissection, Cystic Artery, Cystic Duct, Gallbladder, and Tool (Murali et al., 2023). Endoscapes-BBox201 is a strict subset of Endoscapes-CVS201 at each split (Murali et al., 2023).
Segmentation is available in two forms. The released Endoscapes-Seg50 comprises 50 of the 201 videos and 493 segmentation-annotated frames, while a larger Endoscapes-Seg201 with 1,933 segmentation frames is discussed as a ceiling setting in benchmark experiments (Murali et al., 2023). These segmentation masks cover the same six CVS-relevant structures used for detection. Earlier segmentation work on Endoscapes used a richer 29-class annotation ontology before collapsing to 7 effective classes for experiments, including Background, Gallbladder, Cystic Duct, Cystic Artery, Cystic Plate, HC Triangle, and Instruments (Alapatt et al., 2021).
This alignment across label granularities is methodologically significant. It enables direct comparison between image-only CVS models, object-layout models, segmentation-based graph methods, and mixed-supervision pipelines within a shared 201-case cohort (Murali et al., 2023). It also explains why Endoscapes-CVS201 has been attractive for studies that ask whether CVS can be recognized from image-level labels alone, or whether explicit object and region structure is required.
The same underlying cohort has also been extended with scene-graph supervision. Endoscapes-SG201 is a scene-graph-ready extension of Endoscapes-BBox201 with refined tool boxes, six tool subclasses, six action classes, and hand-identity labels; it is then used to pre-train graph encoders that are fine-tuned on Endoscapes-CVS201 for downstream CVS assessment (Shin et al., 21 Jul 2025). This establishes Endoscapes-CVS201 as the downstream clinical target against which richer structural pretraining is evaluated.
5. Benchmark tasks, metrics, and representative results
The standard Endoscapes-CVS201 task is frame-level prediction of the three binary CVS criteria. The official report evaluates criterion-wise Average Precision and their mean,
along with criterion-wise balanced accuracy and its mean, motivated by strong class imbalance (Murali et al., 2023). The report notes that frame-level overall CVS rates are only around 5–8% across splits, whereas video-level overall CVS is substantially higher at 28–37%, illustrating why plain accuracy would be misleading (Murali et al., 2023).
In the CVS201-only regime, the official baseline results are 51.5 CVS mAP and 63.8 balanced accuracy for ResNet50, 52.1 and 64.1 for ResNet50 with reconstruction, and 57.4 and 66.7 for ResNet50-MoCov2 (Murali et al., 2023). When additional box or segmentation supervision is introduced, performance increases substantially. For BBox201 + CVS201, LG-CVS reaches 63.3 mAP and 74.9 balanced accuracy, while the spatiotemporal SV2LSTG reaches 65.3 mAP and 70.8 balanced accuracy. For Seg201 + CVS201, the best single-frame result reported is LG-CVS at 67.3 mAP and 79.8 balanced accuracy, and the best spatiotemporal result is SV2LSTG at 69.7 mAP and 82.4 balanced accuracy (Murali et al., 2023).
Later work uses Endoscapes-CVS201 to probe different representation regimes. In a multi-modal vision-language study, zero-shot evaluation of existing surgical VLPs on the Endoscapes-CVS201 test set gives 26.18 mAP for SurgVLP, 28.46 for HecVL, 26.64 for PeskaVLP, and 19 for a random baseline, showing a substantial gap between generalist pretraining and fine-grained CVS recognition (Baby et al., 7 Jul 2025). The same paper proposes CVS-AdaptNet, a multi-label adaptation strategy with criterion-specific positive and negative prompts; when adapting PeskaVLP on Endoscapes-CVS201, it reports 57.6 mAP, compared with 51.5 for the ResNet50 image-only baseline and 67.3 for the segmentation-based LG-CVS result used as context (Baby et al., 7 Jul 2025).
Graph-pretraining studies use the dataset differently. In “Towards Holistic Surgical Scene Graph,” SSG-Com is pre-trained on Endoscapes-SG201 and then fine-tuned on Endoscapes-CVS201, achieving 64.6 mAP and improving over LG-CVS at 63.2 in that two-stage setting (Shin et al., 21 Jul 2025). This suggests that Endoscapes-CVS201 functions not only as a standalone benchmark but also as a transfer target for representations learned from anatomy, tools, actions, and hand identity.
6. Ambiguity, derivatives, and broader significance
Several limitations of Endoscapes-CVS201 are explicit in the literature. The annotations are difficult and subjective: later work highlights low inter-annotator agreement, citing Cohen’s among three surgical experts (Baby et al., 7 Jul 2025). The labels are sparse in time, the positive class is rare at the frame level, and the dataset availability is limited enough that recent multi-modal work uses this single dataset for both training and evaluation (Murali et al., 2023, Baby et al., 7 Jul 2025). These characteristics have influenced method design, including the use of inverse-frequency weighting in official baselines and KL-divergence-based batch-contrastive objectives in later multi-modal formulations (Murali et al., 2023, Baby et al., 7 Jul 2025).
At the same time, the dataset has proved extensible. SurgTEMP repurposes Endoscapes CVS labels into an open-ended VQA subset, Endoscapes-VQA, by segmenting videos into fixed 5-second intervals, taking the central frame as keyframe, and generating three question-answer pairs per segment, one for each CVS criterion (Li et al., 31 Mar 2026). The resulting Endoscapes-VQA subset contains 1,812 segments and 5,436 QA pairs, reframing binary CVS assessment as text-grounded explanatory reasoning while preserving the original per-criterion supervision structure (Li et al., 31 Mar 2026). This reuse indicates that Endoscapes-CVS201 can serve both discriminative benchmarking and language-conditioned clinical explanation tasks.
The dataset is publicly distributed through the Endoscapes repository, together with official splits, benchmark code, and model checkpoints for several baselines (Murali et al., 2023). Its broader significance lies in the combination of three properties that are rarely co-located in surgical AI benchmarks: formalized expert CVS labels, companion anatomy/tool annotations at box and mask level, and standardized video-level splits for fair comparison. As a result, Endoscapes-CVS201 has become a common reference point for image-only classifiers, graph-based spatial reasoning, multi-modal foundation-model adaptation, and VQA-style reformulations of surgical safety assessment (Murali et al., 2023, Baby et al., 7 Jul 2025, Shin et al., 21 Jul 2025, Li et al., 31 Mar 2026).