EndoGlacier: A Data Integration Paradigm

• EndoGlacier is a conceptual framework that orchestrates heterogeneous data from surgical workflows and glacier studies to yield integrated, uncertainty-aware outputs.
• It underpins surgical video management with multi-annotator consensus while coupling glacier modeling with real-world hazard assessments.
• The framework extends to remote sensing, ground-based observatories, and diffuse optics, facilitating reconstruction of glacier geometry and internal processes.

EndoGlacier is not a single uniformly defined object across the supplied literature. In one explicit usage, it is the named software infrastructure behind the SAGES Critical View of Safety Challenge: “a Python-based framework for managing large-scale surgical data flows” and, more broadly, “a reproducible, automation-driven framework for coordinating global video sourcing, expert training, multi-annotator workflows, and quality control” (Alapatt et al., 21 Sep 2025). In several glacier-oriented syntheses built around other papers, the same label is used more loosely for end-to-end systems that infer internal glacier state, reconstruct glacier geometry, observe endoglacial processes, or operationalize glacier monitoring from heterogeneous measurements. A plausible unifying interpretation is that EndoGlacier denotes an architectural pattern centered on hidden-state inference, orchestration of heterogeneous observations, and uncertainty-aware downstream use.

1. Scope, nomenclature, and recurring meanings

In the supplied material, the term appears in two distinct ways. First, it is an explicit proper name for a surgical-data orchestration framework (Alapatt et al., 21 Sep 2025). Second, glacier-science syntheses use “EndoGlacier” as an inferred label for integrated frameworks concerned with internal glacier properties, endoglacial observatories, or end-to-end monitoring pipelines (Corcoran et al., 2022, Saintenoy et al., 2013, Gimbert et al., 2020, Allgaier et al., 2021, Vachier et al., 2022, Wu et al., 2023, Maslov et al., 2024, Goger et al., 2021).

Usage in the supplied literature	Status	Core role
SAGES CVS Challenge backbone	Explicit named system	Orchestrates surgical video, annotators, and QC
Glacier state-estimation framework	Inferred usage	Couples observations to internal glacier dynamics
Endoglacial observatory	Inferred usage	Resolves basal, englacial, and hydrological processes
Diffuse-optical probing system	Inferred usage	Recovers thickness and optical properties
Global glacier-monitoring pipeline	Inferred usage	Produces operational outlines and confidence maps

A recurrent misconception would be to treat EndoGlacier as an already standardized glacier-science platform. The supplied literature suggests the opposite. Only the surgical instance is an explicitly named implementation. The glacier-related uses are conceptual extensions: they describe what an EndoGlacier-type system would do, not a single canonical software package.

2. EndoGlacier as an explicit orchestration framework in surgical AI

The clearest definition appears in the SAGES Critical View of Safety Challenge, where EndoGlacier is the orchestration and automation framework that made the benchmark feasible at global scale (Alapatt et al., 21 Sep 2025). It is not a single model or annotation GUI. Instead, it integrates three coordinated streams: annotator management, video management, and orchestration and automation. The framework interacts with MOSaiC, the web-based annotation platform, and sits upstream of the Grand Challenge evaluation environment.

Its purpose was to address concrete bottlenecks: 1,419 donated videos were curated into 1,000 qualified 90-second clips; each clip was labeled by three experts; recruitment and annotation spanned 54 institutions in 24 countries over three years; and the benchmark had to preserve subjectivity while still yielding usable ground truth. EndoGlacier automated stage-to-stage progression, participant status updates, email reminders, error logging, and error handling, while preserving diversity of annotator opinions and preventing any single annotator from dominating the dataset.

The video-management stream performed de-identification, eligibility screening, metadata preannotation, and clip construction. Eligibility required minimally invasive cholecystectomy, a continuous 90-second segment before clipping of cystic duct or artery with clear operative field, and exclusion of bailout procedures or incomplete videos. Objective metadata were dual-annotated and adjudicated until two consecutive raters agreed across all criteria. The output was a final pool of 1,000 qualified 90-second clips with metadata including surgical platform and adjunct imaging.

The annotator-management stream handled recruitment, training, protocol alignment, and competency validation. Of 106 potential annotators, 71 met basic clinical criteria, 67 took the exam, 27 passed the threshold of at least 75% correspondence with expert ratings, and the final pool consisted of 20 qualified expert annotators. EndoGlacier then assigned each qualified clip to three independent qualified annotators in bi-weekly buckets of 20 videos per annotator, using MOSaiC APIs to enforce blinding, pacing, and completion tracking.

The resulting dataset was structurally unusual because it preserved multi-annotator uncertainty. For hard labels in Subchallenge A, the ground truth was majority vote,

$$y = \operatorname{mode}(\{l_i\}), \quad i \in \{1,2,3\}.$$

For calibration in Subchallenge B, the benchmark used a confidence-aware soft target,

$$y = \frac{1}{3}\sum_{i=1}^{3}\left(0.5 + (l_i - 0.5)\cdot c_i\right),$$

with evaluation by Brier Score. This structure depended directly on EndoGlacier’s ability to collect three labels plus confidence for every clip. The challenge outcomes reached up to a 17% relative gain in assessment performance, over 80% reduction in calibration error, and a 17% relative improvement in robustness over the state of the art, which underscores EndoGlacier’s role as infrastructure for uncertainty-aware benchmarking rather than as a predictive model in its own right (Alapatt et al., 21 Sep 2025).

3. Glacier-state inference and subsurface reconstruction

Glacier-oriented uses of EndoGlacier center on inferring unobserved internal state from external measurements. A synthesis built around “Ensemble Kalman Filtering for Glacier Modeling” explicitly frames EndoGlacier as an end-to-end glacier modeling framework that tries to infer internal glacier properties and evolution from external observations (Corcoran et al., 2022). In that formulation, a two-stage marine-terminating glacier box model evolves average glacier thickness $H(t)$ and glacier length $L(t)$, while an Ensemble Kalman Filter corrects state and parameter errors through repeated forecast–analysis cycles. The study reports that ensembles of 7–10 members already give near-optimal reduction in mean square difference; observations every 19 years before 1900 and yearly after 1950 were sufficient in the simplified setting; and modern observations can correct deviations introduced by sparse pre-satellite data. The same workflow was propagated through sea-level and ADCIRC storm-surge calculations, which suggests an EndoGlacier-type system in which hidden glacier state is not only estimated but also linked to downstream hazard models.

A complementary geometry-first interpretation appears in the Ground-penetrating Radar study of Austre Lovénbreen (Saintenoy et al., 2013). There, the paper does not define EndoGlacier explicitly, but the supplied synthesis positions the dense GPR-derived geometry as a direct data foundation for an EndoGlacier framework focused on internal and basal processes. The field campaign produced 67,542 georeferenced ice-thickness points, approximately 14,683 points per km$^2$, interpolated on a 10 m grid. Snow thickness was derived from 42 drillings and corrected using $v_{\text{snow}} = 0.22~\text{m/ns}$; ice thickness conversion used $v_{\text{ice}} = 0.17~\text{m/ns}$. The resulting glacier-wide estimates were an averaged ice thickness of 76 m, a maximum depth of 164 m, and a volume of $0.3487 \pm 0.041~\text{km}^3$ with a relative error of 11.9%. Bedrock topography followed directly from

$$z_{\text{bedrock}}(x,y) = z_{\text{surface}}(x,y) - h_{\text{ice}}(x,y).$$

Taken together, these two lines of work suggest two complementary meanings of EndoGlacier in glaciology. One is dynamical and assimilation-based: estimate hidden state through filtering. The other is structural and geophysical: reconstruct the 3D geometry required by thermomechanical or hydrological models. A plausible implication is that a mature EndoGlacier system would couple both.

4. Endoglacial observatories, biological transport, and glacier microclimate

A more process-resolved interpretation is provided by the RESOLVE experiment, which explicitly describes a dense, multi-instrument “EndoGlacier observatory” on Argentière Glacier (Gimbert et al., 2020). The deployment used 98 autonomous 3-component Fairfield ZLand nodes on the glacier surface, one additional 3-component borehole seismometer at 5 m depth, continuous recording for 35 days in early spring 2018, complementary GPR, drone imagery, GNSS positioning, direct basal sliding measurements, and subglacial water discharge. The dense array geometry, with approximate spacing of ~40 m in the flow direction and ~50 m across-flow, was designed to resolve high-frequency cryoseismology in the 4–50 Hz band. The experiment located basal stick-slip clusters at depths between ~80 and 285 m, mapped seismic tremor associated with subglacial water flow, and related near-surface seismic velocities to crevasse fields. In this usage, EndoGlacier denotes a multi-physics observational system in which internal friction, hydrology, and damage are jointly resolved.

The biological extension is more explicitly inferential. A synthesis based on “Biolocomotion and premelting in ice” treats EndoGlacier as an endoglacial ecosystem or instrument concept built around micro-organisms moving through thin liquid films in ice (Vachier et al., 2022). The key mechanism is interfacial premelting, with film thickness

$d = \frac{R_g T_m^2 N}{\rho_l q_m \Delta T},$

where $$y = \frac{1}{3}\sum_{i=1}^{3}\left(0.5 + (l_i - 0.5)\cdot c_i\right),$$0 includes a biological enhancement factor $$y = \frac{1}{3}\sum_{i=1}^{3}\left(0.5 + (l_i - 0.5)\cdot c_i\right),$$1 capturing EPS or AFP effects. Thermal regelation and chemotaxis then compete in setting trajectories, while the long-time macroscopic transport is described by a Fokker–Planck equation obtained from active Ornstein–Uhlenbeck dynamics. The synthesis emphasizes the ratio $$y = \frac{1}{3}\sum_{i=1}^{3}\left(0.5 + (l_i - 0.5)\cdot c_i\right),$$2 as the control parameter determining whether chemotaxis or premelting-enhanced diffusion dominates. This suggests an EndoGlacier concept in which internal glacier habitability, transport, and biosignatures are modeled as coupled physical–biological processes rather than as static inclusions in ice.

A third coupling layer involves the glacier boundary layer. Large-eddy simulations over Hintereisferner at $$y = \frac{1}{3}\sum_{i=1}^{3}\left(0.5 + (l_i - 0.5)\cdot c_i\right),$$3 m show that glacier microclimate can either persist or be eroded depending on synoptic flow direction (Goger et al., 2021). Under southwesterly airflow, the down-glacier wind is supported by the synoptic direction and a stable boundary layer forms over the ice surface. Under northwesterly airflow, a cross-glacier valley flow and a breaking gravity wave lead to strong turbulent mixing and erosion of the glacier boundary layer. The study’s heat-budget and TKE analyses indicate that exchange mechanisms are fully three-dimensional rather than only vertical. In an extended EndoGlacier reading, the “inside” of the glacier system therefore includes not only basal and englacial processes but also the atmospheric shell that controls turbulent melt fluxes.

5. Diffuse optics and remote-sensing operationalization

Another inferred EndoGlacier modality is low-cost optical probing of glacier interiors. “Diffuse optics for glaciology” proposes a surface-based system in which short optical pulses are injected into glacier ice treated as a turbid, diffusive medium (Allgaier et al., 2021). In the diffusion approximation,

$$y = \frac{1}{3}\sum_{i=1}^{3}\left(0.5 + (l_i - 0.5)\cdot c_i\right),$$4

and, for forward-peaked scattering, the diffusion coefficient reduces to $$y = \frac{1}{3}\sum_{i=1}^{3}\left(0.5 + (l_i - 0.5)\cdot c_i\right),$$5. The paper uses Monte Carlo simulations with Mie-like scattering, Henyey–Greenstein phase functions with $$y = \frac{1}{3}\sum_{i=1}^{3}\left(0.5 + (l_i - 0.5)\cdot c_i\right),$$6, Fresnel reflection at the surface, and mixed boundary conditions. The results suggest that effective scattering length and absorption length can be recovered with relative errors below ~20% at suitable source–detector separations, while glacier thickness up to ~25 m can be retrieved under favorable reflective-bottom conditions. The supplied synthesis characterizes this as an EndoGlacier-type system because it extracts internal optical properties and geometry from non-invasive surface measurements.

At the glacier surface and inventory scale, the literature shifts from internal probing to operational mapping. AMD-HookNet formulates glacier calving-front extraction from SAR as multi-class zone segmentation followed by boundary extraction, using a dual-branch U-Net with attention, multi-hooking, and deep supervision (Wu et al., 2023). On the CaFFe benchmark, the method reduces mean distance error from $$y = \frac{1}{3}\sum_{i=1}^{3}\left(0.5 + (l_i - 0.5)\cdot c_i\right),$$7 m to $$y = \frac{1}{3}\sum_{i=1}^{3}\left(0.5 + (l_i - 0.5)\cdot c_i\right),$$8 m, a 42% improvement over the baseline. In an EndoGlacier framing, this is an external manifestation of internal or terminus dynamics operationalized for automated monitoring.

GlaViTU then extends the operational layer to global glacier mapping (Maslov et al., 2024). The model combines a transformer subnet, a convolutional U-Net-like subnet, squeeze-and-excitation fusion, and region or coordinate encoding. The released benchmark covers about 9% of glaciers worldwide. The best strategy, region encoding with bias optimization, achieves IoU $$y = \frac{1}{3}\sum_{i=1}^{3}\left(0.5 + (l_i - 0.5)\cdot c_i\right),$$9 on previously unobserved images in most cases, drops to $H(t)$0 for debris-rich areas, and increases to $H(t)$1 for clean-ice dominated regions. Adding synthetic aperture radar data, namely backscatter and interferometric coherence, increases the accuracy in all regions where available. Confidence calibration is central: Expected Calibration Error drops from $H(t)$2 to $H(t)$3 for plain softmax and from $H(t)$4 to $H(t)$5 for Monte-Carlo dropout after calibration. This suggests an EndoGlacier pipeline in which internal-state inference and surface delineation are coupled to confidence-aware global inventory production.

6. Architectural motifs, limitations, and future directions

Across these uses, EndoGlacier is best understood as an architectural pattern rather than a single scientific object. The explicit surgical framework and the inferred glacier systems share a recognizable set of motifs: orchestration of heterogeneous inputs, separation of concerns between acquisition and analysis, explicit treatment of uncertainty, and downstream evaluation under clinically or geophysically meaningful heterogeneity (Alapatt et al., 21 Sep 2025, Corcoran et al., 2022, Gimbert et al., 2020, Maslov et al., 2024).

Several limitations recur. In the surgical instance, confidence was captured per clip rather than per frame, and the authors leave finer-grained uncertainty modeling for future work (Alapatt et al., 21 Sep 2025). In the glacier-assimilation prototype, the two-stage model is intentionally simple, uses synthetic observations, and would require more complicated glacier models and real observation data for broader validity; efficient filters such as SEIK are explicitly suggested as future work (Corcoran et al., 2022). In the GPR geometry case, interpolation error dominates the uncertainty budget, especially away from survey tracks (Saintenoy et al., 2013). In the dense seismic observatory, future advances are tied to denser detection catalogues, double-difference relocation, and stronger process-level links between tremor, slip, and hydrology (Gimbert et al., 2020). In the optical probing formulation, thickness retrieval is sensitive to lower-boundary reflectivity and to assumptions of homogeneous scattering (Allgaier et al., 2021). In operational mapping, calving fronts, ice mélange, debris-covered tongues, and shadowed ice remain the main failure modes, even when attention mechanisms, InSAR coherence, and calibrated confidence are used (Wu et al., 2023, Maslov et al., 2024).

Future directions are correspondingly convergent. The surgical literature proposes reuse of EndoGlacier’s orchestration pattern for new procedures and quality metrics (Alapatt et al., 21 Sep 2025). Glacier modeling papers point toward higher-fidelity data assimilation, probabilistic handling of internal geometry, and tighter coupling to hazard models (Corcoran et al., 2022, Saintenoy et al., 2013). Observatory-style work points toward real-time endoglacial monitoring with dense arrays and integrated GPR, GNSS, and hydrology (Gimbert et al., 2020). Biosignature and microclimate studies suggest that an expanded EndoGlacier concept may need to represent premelting, chemotaxis, katabatic flows, and gravity-wave-driven boundary-layer erosion within a single coupled system (Vachier et al., 2022, Goger et al., 2021). Remote-sensing work points toward globally scalable, uncertainty-aware inventories that can be specialized for difficult regimes such as calving fronts and debris-covered glaciers (Wu et al., 2023, Maslov et al., 2024).

In that sense, EndoGlacier is less a fixed platform than a family resemblance among systems designed to make hidden structure observable. In surgery, it coordinates expert judgment and data flow at global scale. In glacier science, the supplied literature uses it as a label for frameworks that recover geometry, dynamics, biology, acoustics, optics, and mapped extent from incomplete measurements. The common principle is not domain-specific nomenclature but the systematic conversion of heterogeneous observations into reproducible, uncertainty-aware representations of otherwise inaccessible internal state.