Cohort Co-Registration: Sharing Liver Deformation Across Patients When Each Case Has Only Weak Evidence

During laparoscopic liver surgery, the surgeon would like the preoperative CT model of the liver overlaid on the live video. The liver, however, bends and compresses once the abdomen is inflated with gas and instruments push on it. Current systems estimate this non-rigid deformation one patient at a time, from whatever the camera shows: a liver outline and a few visible ridges. That evidence is thin, and it gets worse when the video comes from a different hospital's equipment. We propose to register a whole cohort of patients jointly. Each patient's deformation is pulled toward the deformations of anatomically similar patients, with similarity measured on the preoperative liver shape rather than on video appearance. Weak evidence from many cases is then pooled instead of being wasted case by case.

This is a population graph, meaning a graph whose nodes are patients and whose edges link similar patients. Here it is used for a reason that survives the strongest criticism of such graphs. Are Population Graphs Really as Powerful as Believed? showed that on five medical benchmarks a Random Forest on concatenated features matches or beats graph neural networks: 0.9474 versus 0.9404 accuracy on TADPOLE, and 0.7046 versus 0.695 on ABIDE. The authors conclude that edges help only when they carry information "that cannot be trivially added to node features." In our setting the quantity passed along an edge is a neighbor's jointly estimated deformation. That is an inferred output, not an input feature, and it is refined together with the patient's own. The proposal is built to test whether this relational coupling actually helps, and in which regime.

Core Claim

Joint estimation should reduce registration error compared with per-patient registration whenever per-patient evidence is weak. Examples are few annotated landmarks, degraded video, or a new acquisition center. The coupling must act in a space where deformations can be averaged sensibly. It must also use edges built from geometry that is insensitive to scanners. The advantage should shrink to zero as per-patient evidence becomes rich, and it should vanish if the edges are shuffled. If neither prediction holds, the graph is decoration.

Concretely, each liver deformation is written as $\phi_i = \exp(v_i)$. Here $v_i$ is a stationary velocity field, a smooth vector field whose integration yields a fold-free deformation. We represent $v_i$ in a common liver template frame, which makes velocities from different patients directly comparable and averageable. This log-Euclidean linearization comes from MORPH-LER: Log-Euclidean Regularization for Population-Aware Image Registration, which uses it to regularize brain MRI registration toward population-wide deformation modes. MORPH-LER learns one global population prior from fully supervised static volumes. We instead couple individual cases through a sparse graph and solve them jointly from weak intraoperative evidence. The cohort objective is

$$ \min_{\{v_i\}} \; \sum_i \mathcal{L}_i\big(\exp(v_i)\big) \;+\; \lambda \sum_{(i,j)} w_{ij}\, \lVert v_i - v_j \rVert^2 . $$

In this objective, $\mathcal{L}_i$ is patient $i$'s weak evidence loss. It is made of three terms taken from Preoperative-to-intraoperative Liver Registration for Laparoscopic Surgery via Latent-Grounded Correspondence Constraints (Land-Reg): silhouette agreement with an automatic SAM2 liver mask, reprojection of whatever landmark curves are available, and a local-isometry penalty that discourages non-physical stretching. The weight $w_{ij}$ is the edge between patients $i$ and $j$, computed from preoperative liver geometry. The coefficient $\lambda$ sets the coupling strength.

A uniform, fully connected graph reduces this objective to shrinkage toward the cohort mean, which is essentially the MORPH-LER-style global prior. Setting $\lambda = 0$ recovers per-case Land-Reg. Both are therefore exact special cases and serve as natural controls.

Research Program

Building the template frame. We first build a liver template and map every patient's preoperative segmentation to it. Velocity fields can then be transported into the template frame. Transport errors are a known approximation. We will measure them by round-trip consistency on held-out shapes.

Designing the edges. Edges come from geometric descriptors of the segmented preoperative liver: template-frame shape coefficients, lobe volume ratios, and the curvature of the inferior ridge. The rationale draws on AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation, which reports drops of over 20% mDice on unseen MRI scanners. That result suggests intensity-derived embeddings will encode the scanner. Organ geometry extracted from a segmentation should depend less on it. We test this rather than assume it, by building rival graphs from CNN intensity embeddings.

Handling uncertain edges. As an extension, edge weights can be treated as uncertain, following the variational edge formulation of Edge-variational Graph Convolutional Networks for Uncertainty-aware Disease Prediction. Here the uncertainty would be fitted to how well neighbors' deformations explain each other's evidence, rather than to classification labels.

Extracting intraoperative evidence. Landmark curves come from a topology-preserving detector. We use Topology-Constrained Learning for Efficient Laparoscopic Liver Landmark Detection (TopoNet, 65.19% Dice and 28.07 px ASSD on L3D), together with a controlled number of manual curves. Varying that number gives the annotation-budget axis. A secondary question is which patients to annotate densely. We will compare random selection against choosing graph-central patients, whose solved deformations propagate furthest.

Data. Real data is scarce. P2ILF has only 9 patients and 167 images, too few to study graph structure alone. We therefore add a simulated cohort. We take AMOS liver segmentations (500 CT scans from 8 scanners and 2 centers), apply biomechanically simulated gas-inflation deformations with varied tissue stiffness, and render silhouettes and ridge curves from sampled laparoscope poses. The simulation gives ground-truth deformations at population scale and a built-in preoperative domain shift through the scanner and center splits. P2ILF remains the real-data test, used with leave-one-patient-out evaluation.

Baselines

Rigid alignment only. The do-nothing reference is rigid alignment, which Land-Reg reports at 69.21% Dice on P2ILF.

Per-case non-rigid registration. The strongest existing per-case method is Land-Reg's non-rigid stage. It reports 75.38% Dice and a mean target registration error of 76.46 pixels on P2ILF, measured on the 9-patient cohort. We will rerun Land-Reg ourselves with the identical evidence loss, so that the only difference from our method is $\lambda$. Beating it means a lower leave-one-patient-out error on P2ILF with matched landmark budgets, consistent across patients under a paired test, plus a clear gain on simulated held-out cases.

Global mean prior. The MORPH-LER-style global prior shrinks every patient toward the cohort mean velocity, or toward a low-rank principal-component model of velocities. We implement it as the uniform-graph special case with $\lambda$ tuned on the same validation split. Beating it is the main evidence that which patients are coupled matters.

Anatomy-to-deformation regressor. This is the Müller-style baseline. Gradient-boosted trees, plus a small MLP, predict velocity principal-component coefficients from the same preoperative shape descriptors the graph uses. The prediction serves as a per-case prior. The regressor is trained only on the cases that have dense annotations under each budget. It has the same features and the same tuning budget as our method, and it is the most dangerous competitor.

Edge ablations. Two ablations isolate the edges. In the first we shuffle the edges while keeping the degree distribution. In the second we build edges from CNN intensity embeddings of the CT. The proposal survives only if geometry edges beat shuffled edges, and only if they beat intensity edges on cross-scanner splits.

Why not just…?

Why not just use a single population-mean deformation prior? If deformation under gas inflation is roughly the same for every liver, a global prior captures most of the benefit, and the graph adds nothing. The uniform-graph baseline tests this exactly. In simulation we can vary how much deformation depends on anatomy and stiffness. Our method should win only when that dependence is heterogeneous. Showing where the crossover lies is itself a result.

Why not just regress the deformation from preoperative anatomy? This is the real threat. With enough densely annotated cases, a supervised regressor from shape descriptors to deformation coefficients may equal the graph, just as Are Population Graphs Really as Powerful as Believed? found for classification. Our claim is limited to the low-annotation, transductive regime. There, the regressor has few labeled targets, while joint inference also draws on weakly evidenced, unlabeled cases. The annotation-budget sweep decides the question. If the regressor matches us at every budget, we will report that the relational component is unnecessary and that a population prior suffices.

Why not just strengthen per-case regularization or rely on foundation models? Stronger isometry penalties or better SAM2 and depth priors improve each case's evidence, but they cannot supply information the camera never saw, such as the posterior surface. The evidence that per-case foundation priors break under degradation comes from endoscopic reconstruction: EndoMD-SLAM: Endoscopic Gaussian Splatting SLAM under Optical Degradation with Memory and Static-Transient Decomposition reports around 35 mm trajectory error for standard Gaussian-splatting SLAM on degraded colonoscopy video. That is a different task from registration, so it motivates our concern rather than settling it. We test the concern directly by degrading P2ILF and simulated frames with smoke, blur, and specular corruption, and checking whether coupling's benefit grows as the evidence degrades.

Why not just retrieve the nearest patient and copy its deformation? One-shot retrieval is a single message-passing step with no joint refinement. It inherits the neighbor's errors. We run it as an ablation by fixing neighbors at their per-case solutions, which separates the value of joint inference from the value of the neighborhood itself.

Proposed Experiments

Experiment 1: simulated evidence and budget sweep. In the AMOS-based simulated cohort, train on one center's scanners and test on the other's. We sweep the number of visible landmark curves from zero to full and the level of optical degradation. Metrics are 3D target registration error against simulated ground truth, the fraction of folded voxels (negative Jacobian determinant), and silhouette Dice. The decisive check is a dose-response pattern. The gain of geometry-edge coupling over per-case Land-Reg should grow as evidence weakens. It should exceed both the global prior and the regressor at low budgets. Shuffled edges should erase it.

Experiment 2: real-data test on P2ILF. With leave-one-patient-out evaluation, we report 2D target registration error in pixels and Dice alongside the published 76.46 px and 75.38%. The cohort graph for each held-out patient includes the other eight real patients plus the simulated cohort, so this experiment also measures the sim-to-real gap. With only 9 real patients this experiment has limited statistical power. We treat it as a consistency check on the simulated findings, not as the main evidence.

Experiment 3: edge robustness across scanners. We compare graphs built from geometry against graphs built from intensity embeddings, with test CTs drawn from unseen AMOS scanners. If intensity edges degrade on unseen scanners while geometry edges do not, that supports the domain-invariance argument. If both behave the same, the argument fails and we will say so.