Soft tissue holds its shape because of the extracellular matrix, the mesh of collagen and other proteins between cells. Enzymes called matrix metalloproteinases (MMPs, which cut collagen) weaken that mesh. A large body of biology studies drugs, microRNAs and endogenous proteins that suppress these enzymes in order to "stabilize" the matrix. Those studies almost always measure success at a single endpoint, using destructive assays: zymography (a gel that shows enzyme activity), Western blots, or counting the cells that cross a membrane coated with Matrigel (a basement-membrane gel). None of these show where or when a piece of tissue actually got mechanically weaker. Surgical computer vision has the opposite gap. Dynamic 3D reconstruction methods now track deforming tissue from endoscope video at more than 50 frames per second, but they only move visual primitives to match pixels. They estimate nothing about how stiff the tissue is.
We propose to join these two lines with an inverse problem. A camera films a matrix sample, or a piece of tissue, while a sensorized probe pushes it with a known force. We fit a deforming 3D Gaussian-splat reconstruction whose primitives carry a stiffness value that varies over space and time and must satisfy mechanical equilibrium. The output is a time-lapse stiffness map. On that map, enzymatic softening and its suppression by MMP inhibitors become visible as they happen, without destroying the sample. The vision community gains a constitutive-parameter estimator that can be checked against physical ground truth. Matrix biology gains a live, spatial readout of "stabilization" in place of a bulk endpoint.
Core Claim
Jointly fitting appearance, geometry and a heterogeneous, time-varying stiffness field, constrained by quasi-static equilibrium under measured probe force, can recover a matrix region's local softening relative to untreated matrix. It can also recover how fast that softening proceeds. In particular, it should separate enzyme-treated, inhibitor-protected and untreated regions of the same sample. A two-stage pipeline (track first, then fit a finite element model) or homogeneous physics regularization should not be able to do this. The new element is not "physics in Gaussian splatting," which already exists. The new element is a spatially and temporally varying material parameter that is identifiable from video plus force, and that is validated against independent mechanical and biochemical measurements of matrix degradation.
Relation to Prior Work
The method draws on two existing pieces. Physics-Informed Deformable Gaussian Splatting (PIDG) treats each Gaussian as a material point and penalizes violations of the Cauchy momentum equation. It assumes homogeneous, static elastoplastic material and is evaluated on smoke, cloth and toys, judged by rendering quality. We keep its "Gaussians as material points" idea and change three things. The material parameter becomes a learned field over space and time. The dynamics simplify to quasi-static equilibrium, which suits slow probing. The measured probe force fixes the absolute stiffness scale, which video alone cannot do. Efficient EndoNeRF reconstruction and its application for data-driven surgical simulation exports surgical reconstructions into finite element and material point method solvers. Its elastic parameters are fixed and synthetic, and the physics runs only downstream, so reconstruction never learns from mechanics. Its export pipeline, including closing open surfaces into volumes, becomes our two-stage baseline.
The kinematic reconstruction line motivates the gap. EndoGS, Endo-4DGS, HFGS, and the plane-factorized Forplane and LerPlane all fit deformation to photometric, depth or flow losses with no material model. This is fine for rendering but silent about stiffness. Endo-4DGS also supplies a constraint: monocular depth priors carry scale ambiguity. We therefore use stereo throughout and treat monocular input as a stress test. The SLAM systems EndoGSLAM and EndoMD-SLAM assume a rigid or static anatomy field. Slow enzymatic creep would be absorbed as tracking error. This is why our first setting uses a fixed, calibrated camera, which separates mechanics from pose estimation.
On the biology side, the matrix-stabilization literature defines what we want to see. Gelatinases MMP-2 and MMP-9 degrade basement-membrane collagen. Their suppression blocks passage through Matrigel (Ligularia fischeri inhibits..., Broussonetia kazinoki modulates...). Endogenous inhibitors such as RECK restrain MMP-2, MMP-9 and MT1-MMP (RECK as a Potential Crucial Molecule for the Targeted Treatment of Sepsis). The RECK review itself notes that evidence in dynamic organ models is scarce and is mostly extrapolated from static in vitro work. Tissue Engineering of Endothelial Cells and the Immune Response shows that a 3D collagen matrix changes the phenotype of the cells embedded in it. This motivates a later cell-laden experiment in which matrix mechanics are the quantity of interest. We do not model enzyme biochemistry inside the vision model. Enzyme activity is the experimental manipulation, and stiffness is the measured response.
Research Program
The representation is a set of Gaussians, each with a rest position, appearance attributes and a log-stiffness value. The log-stiffness is read from a low-rank space-time plane grid, in the style of HexPlane and Forplane, so that it is smooth and cheap to store. Under each probe press at time $t$, Gaussian displacements $u$ are predicted. A neo-Hookean stress $\sigma(E,u)$ is computed on a tetrahedral mesh that we refit periodically to the Gaussian centres; neo-Hookean is a standard nonlinear elastic model that remains valid at large strains. The training objective is
$$\mathcal{L} = \mathcal{L}_{\text{photo}} + \lambda_d \mathcal{L}_{\text{stereo depth}} + \lambda_m \sum_{i} \left\| \nabla \cdot \sigma(E_i(t), u_i) + f_i \right\|^2 + \lambda_s \|\partial_t \log E\|^2,$$
The symbols are defined as follows. $E_i(t)$ is the stiffness at Gaussian $i$ and time $t$. $f_i$ is the external force: it equals the measured probe load distributed over the contact patch and is zero elsewhere. The third term enforces quasi-static equilibrium. The last term discourages stiffness from flickering between presses, while still letting it drift slowly as enzymes act.
Because elastic equilibrium fixes stiffness only up to the force scale, the measured force is what makes $E$ identifiable in absolute units. Without the force, only relative stiffness maps are meaningful. We will report both and say which applies in each setting. We will also analyze identifiability directly. A region that the probe barely strains carries little information about its stiffness. The model therefore outputs a per-region sensitivity estimate, the inverse Hessian of the equilibrium term with respect to $\log E$. Low-confidence regions are flagged rather than reported.
The experimental platform is deliberately simple. It consists of a calibrated stereo endoscope or stereo camera pair fixed over a dish or tissue holder, and a motorized indenter with a load cell that presses at programmed sites every few minutes for one to three hours.
Baselines
The naive reference fits a single homogeneous modulus per sample from probe force and indentation depth: the textbook bulk measurement. It sets the floor for spatial detection. Any useful method must separate regions the bulk fit cannot.
The strongest kinematic competitor is HFGS, run as a two-stage pipeline. Its tracked displacements feed an offline finite element inversion that solves for the stiffness field, using the open-to-closed meshing of FastEndoNeRF. On the EndoNeRF benchmark, HFGS reports PSNR 38.14 against EndoGS at 36.84, so it is the best reported tracker in the file. No reported result was found on Lacuna for any method's stiffness-recovery accuracy. The mechanical comparison therefore has no published number to beat. It will be decided against our own indentation ground truth.
PIDG is run with a homogeneous, static material, as published. It tests whether physics regularization without a heterogeneous, time-varying parameter suffices.
Two ablations isolate the key ingredients: our model with the stiffness field frozen in time, and our model with spatial variation removed. Separately, we run our model with the force term removed, to measure what video-only relative maps can and cannot deliver.
For fairness, all methods receive identical stereo frames, masks, force logs and mesh resolution. Each gets the same hyperparameter search budget, roughly 20 configurations tuned on held-out calibration phantoms, and the same single-GPU time cap per sequence. Beating the two-stage baseline means statistically lower stiffness-map error against indentation ground truth, together with better detection of treated regions, across samples rather than on one sequence. Rendering quality on EndoNeRF must stay within about 1 dB PSNR of the HFGS figure.
Why not just…?
Why not just use digital image correlation or embedded-bead tracking, the standard tools of gel mechanics? This is the most serious alternative. On flat, speckled or bead-laden gels it may track displacement as well as our reconstruction, and the file contains no evidence otherwise. We therefore include it on speckled flat gels as a displacement baseline, feeding its displacements into the same inversion. If it matches us there, we concede that regime. The remaining claim then rests on curved, unspeckled, wet and specular samples viewed through an endoscope-like optic, which is the regime surgical data comes from. Even then, digital image correlation yields displacement, not stiffness, so it competes with our tracker, not with the joint estimator.
Why not just measure stiffness directly with micro-indentation or rheometry and skip the vision? These methods are the ground truth, which is exactly why they cannot be the method. Rheometry is bulk. Dense micro-indentation mapping is slow, perturbs the sample, and cannot be repeated every few minutes over a whole construct during digestion. We use sparse indentation on sister samples and at endpoints for validation. If sparse indentation alone, interpolated over space, predicts the dense map as well as our method, that would undercut the method. We test this by holding out indentation sites.
Why not just track with an existing 4D splatting model and fit a finite element model afterwards? This is our main baseline, and it may well work on large, high-contrast softened regions. Our argument that it falls short is specific. Tracking errors in low-texture or specular areas, which HFGS and EndoMD-SLAM both document as failure modes, propagate unchecked into the inversion. In the joint model, by contrast, equilibrium can veto implausible displacements. The experiment that settles this compares the two pipelines on stiffness error as a function of surface texture and specularity, under matched compute.
Proposed Experiments
The first study uses gelatin and collagen-I phantoms with known patterned stiffness, made by moulding inclusions of different concentration. It establishes recovery accuracy against indentation maps. The metrics are relative modulus error, Spearman correlation with indentation, and detection area under the ROC curve for the stiff or soft inclusions.
The second study is the core biological test. Each gelatin gel is divided into three zones. One zone receives recombinant MMP-9 or collagenase. A second receives the same enzyme together with a broad MMP inhibitor at several doses. A third receives buffer. The gel is probed every five minutes for two hours. The decisive checks are three. The recovered softening rate must order the zones correctly. It must follow the inhibitor dose-response monotonically. And its ranking across doses must agree with gelatin zymography of sister samples. If the stiffness maps fail to distinguish inhibitor-protected from untreated zones when zymography clearly does, the central claim fails.
The third study repeats the design on ex vivo porcine tissue with local enzyme injection. It is filmed with a stereo laparoscope under realistic specular lighting and evaluated against endpoint indentation and histology. A stretch study tracks compaction and stiffening of endothelial-cell-laden collagen constructs, in the spirit of Methe and Edelman, with and without MMP inhibition.
Finally, the public datasets are used as checks rather than as mechanical ground truth, because they have no force data. These datasets are EndoNeRF, StereoMIS and SCARED. On them we verify that rendering quality does not regress relative to HFGS and EndoGS. We also verify that relative stiffness maps are reproducible across repeated retractions of the same tissue. We make no clinical claim about MMP state from these videos.
References
- Physics-Informed Deformable Gaussian Splatting: Towards Unified Constitutive Laws for Time-Evolving Material Field
- Efficient EndoNeRF reconstruction and its application for data-driven surgical simulation
- EndoGS: Deformable Endoscopic Tissues Reconstruction with Gaussian Splatting
- Endo-4DGS: Endoscopic Monocular Scene Reconstruction with 4D Gaussian Splatting
- HFGS: 4D Gaussian Splatting with Emphasis on Spatial and Temporal High-Frequency Components for Endoscopic Scene Reconstruction
- Efficient Deformable Tissue Reconstruction via Orthogonal Neural Plane
- Neural LerPlane Representations for Fast 4D Reconstruction of Deformable Tissues
- EndoGSLAM: Real-Time Dense Reconstruction and Tracking in Endoscopic Surgeries Using Gaussian Splatting
- EndoMD-SLAM: Endoscopic Gaussian Splatting SLAM under Optical Degradation with Memory and Static-Transient Decomposition
- Ligularia fischeri inhibits endothelial cell proliferation, invasion and tube formation through the inactivation of mitogenic signaling pathways and regulation of vascular endothelial cadherin distribution and matrix metalloproteinase expression
- Broussonetia kazinoki modulates the expression of VEGFR-2 and MMP-2 through the inhibition of ERK, Akt and p70S6K-dependent signaling pathways: Its implication in endothelial cell proliferation, migration and tubular formation
- RECK as a Potential Crucial Molecule for the Targeted Treatment of Sepsis
- Tissue Engineering of Endothelial Cells and the Immune Response