SMART-KNEE — Imageless knee reconstruction from four intraoperative landmarks

Four points, and the rest is inference

Imageless arthroplasty drops the preoperative scan, and with it the patient's anatomy. What the surgeon has instead is a tracked probe and a few minutes. Shape models fill the gap with a population average, which is where axis alignment and implant sizing start to drift.

SMART-KNEE infers the whole bone instead. A topology-aware equivariant graph network completes the landmark set from four anchors; an anatomy-aware GAN grows the surface around them, conditioned on whatever the probe managed to touch.

Where the error lands

Femur first, tibia second. Every figure sits inside the 3 mm clinical threshold.

Dice score: F87.1%; T92.5%
HD95: F5.18; T4.31
Average surface distance: F2.43; T1.70
Normal surface distance: F0.450; T0.728
Mesh to ground truth: F1.97; T2.87
Landmark error: F2.28; T1.65

Complete the landmarks, then grow the bone

Two stages, cascaded. The first is geometric, the second generative.

Topology-aware EGNN

Landmarks are the nodes of a knee graph: four known anchors, eight femoral and seven tibial unknowns. A Kabsch–Umeyama fit of the mean canonical shape sets the initial coordinates, so the network starts from an anatomically plausible configuration rather than noise.

Each node carries its known or unknown status, knee chirality, and geometric descriptors. Four E(3)-equivariant message-passing layers refine both features and 3D positions through learnt gated aggregation, and a decoder MLP applies a clipped coordinate correction in millimetre space. Equivariance means the prediction follows the patient's pose rather than depending on it.

Known anchor nodes in red are mapped through an encoder, edge MLP, node MLP and decoder to a complete set of predicted landmarks in green for both femur and tibia. — Known anchors (red) become the complete landmark set (green) through equivariant message passing.

Anatomy-aware GAN

A PointNet auto-encoder is first trained on complete bones under a symmetric Chamfer distance, then frozen. That fixes a latent geometry manifold on which every plausible knee already lives.

The generator concatenates a patch encoder over the sparse probe cloud with a landmark MLP encoder into a 256-dimensional code. The frozen decoder renders that code as a full point cloud, while a latent discriminator holds the code on the manifold under a least-squares objective. Poisson reconstruction closes the result into a watertight mesh, mirrored for left-sided knees.

Anatomy-aware GAN architecture showing the clean PointNet encoder and decoder, patch and landmark encoders, latent discriminator, conditional generator and frozen decoder. — Ground truth, sparse patch and landmarks feed the two-phase architecture.

What training minimises

L_CD: Symmetric Chamfer distance. Pre-trains the auto-encoder and fixes the geometry manifold.
L_pos + L_edge: Positional L₁ error plus inter-landmark distance deviation, which keeps the predicted landmark set topologically consistent.
L_adv: Least-squares adversarial term in latent space, driving the generated code onto the pretrained manifold.
L_patch + L_lmk: Patch surface coverage and landmark fidelity, anchoring the reconstruction to what the probe actually touched.

L_total = α L_adv + β L_patch + γ L_lmk

Six times closer, and it holds at the hard nodes

361 pairs, split 70 / 15 / 15 by side. Surfaces measured on 39 held-out subjects.

Landmarks

Mean per-node localisation error in millimetres, across all nodes.
Method	Femur	Tibia
VN-EGNN	7.50	7.81
EquiPPIS E(3)	7.91	12.12
SMART-KNEE	1.25	1.69

Grouped bar chart of per-node localisation error for femur and tibia, comparing VN-EGNN, EquiPPIS E(3) and SMART-KNEE against the 3 mm threshold. — Every predicted node falls inside the 3 mm threshold, with the largest gains at nodes furthest from the anchors — the proximal-posterior cortex on the femur and the malleolar landmarks on the tibia.

Surfaces

Reconstruction quality on the 39 held-out test subjects.
Metric	Femur	Tibia
Dice score	87.10%	92.46%
HD95	5.18 mm	4.31 mm
Average surface distance	2.43 mm	1.70 mm
Normal surface distance at 2 mm	0.450	0.728
Mesh to ground truth	1.97 mm	2.87 mm

Four scatter plots of per-sample femur Dice, HD95, ASD and NSD with mean lines. — Femur, per sample.

Four scatter plots of per-sample tibia Dice, HD95, ASD and NSD with mean lines. — Tibia, per sample.

On a printed phantom

Femoral and tibial phantom with predicted landmarks in blue and acquired surface point clouds in colour, probed under optical tracking. — Predicted landmarks in blue, acquired point clouds in colour, on a 3D-printed CT-derived phantom under optical tracking.

The end-to-end pipeline was run under realistic intraoperative conditions on a 3D-printed, CT-derived phantom digitised with an optically tracked probe.

Measure	Femur	Tibia
Landmark RMSE	0.25 mm	0.36 mm
Surface distance	0.38 mm	0.26 mm

Both measures land an order of magnitude inside the clinical threshold, which motivates extending validation to cadaveric specimens and live arthroplasty.

Does it survive an operating theatre?

Anchors digitised on the phantom, then sparse surface sweeps across femur and tibia — the inputs the network actually receives.

Take the data

361 annotated femur and tibia models, landmarks included, on Zenodo. Code on GitHub.

Twelve femoral and eleven tibial landmarks per pair, with the distal and proximal regions used to simulate probe acquisition. Four landmarks per bone act as the known anchors. The femur you can rotate above is one of them, drawn at its released coordinates.

Dataset on Zenodo
github.com/IGRS-Imaging/SMART_KNEE

git clone https://github.com/IGRS-Imaging/SMART_KNEE.git
cd SMART_KNEE
pip install -r requirements.txt

Cite

Pages and DOI to follow once the proceedings are out.

@inproceedings{rajasekar2026smartknee,
  title     = {SMART-KNEE: Sparse-Morphology based Anatomical
               Reconstruction using Topology-Aware Networks for
               Imageless Total Knee Arthroplasty},
  author    = {Rajasekar, Durga and Lakshmi S, Swetha and
               M R, Vishnu and Maik, Vivek and
               Lakshmanan, Manojkumar and Sivaprakasam, Mohanasankar},
  booktitle = {Medical Image Computing and Computer Assisted
               Intervention (MICCAI)},
  year      = {2026},
  publisher = {Springer}
}