Skip to content
Navyashree N
← All work

IIT Madras · HTIC2026

Sub-10 ms bone morphing during surgery, with no preoperative CT — the patient's femur and tibia reconstructed from points swept with a probe on the table.

In progress — paper in preparation, MICCAI 2026 workshop track

Image slot

GPUNeuroMorph — neural signed distance field morphing a femur surface

1600×1000 · pipeline diagram or a render of the morphed femur/tibia surface

Role

Research engineer — algorithm design, CUDA implementation, benchmarking

Year

2026

Stack

  • CUDA
  • Holoscan SDK
  • PyTorch
  • C++
  • Neural SDF
  • Python

The problem

Total knee arthroplasty navigation normally begins with a preoperative CT. The patient is scanned, the bone is segmented, and the surgical plan is built against that model. The scan costs time, money, and radiation dose, and it fixes the plan to anatomy captured days before the incision. The alternative is to build the model in theatre, from points the surgeon sweeps across exposed bone — but that only works if the morphing is fast enough to feel instantaneous and accurate enough to plan a cut against. Those two requirements pull directly against each other, and the usual answer is to give up some of one.

Approach

  1. Represent the bone as a neural signed distance field — a multiresolution hash grid feeding a small MLP — so the surface is continuous and queryable at any resolution rather than frozen into a mesh at authoring time.

  2. Warp in field space using Gaussian radial basis functions rather than displacing vertices. The surface stays watertight under deformations large enough to break a mesh-based approach.

  3. Register the sparse probed points against the statistical model with Coherent Point Drift over hierarchical Gaussian mixtures, keeping the correspondence search well below the quadratic cost of the naive formulation.

  4. Schedule the entire path as a Holoscan SDK operator graph so the intraoperative loop never round-trips to the host — the data stays resident on the GPU from probe input to rendered surface.

Results

GPU latency, full morph

~7 ms

GPU latency, full morph

femur surface error

0.47 mm

femur surface error

tibia surface error

0.33 mm

tibia surface error

preoperative CT scans required

zero

preoperative CT scans required

Internal benchmarks measured on the lab rig. The work is ongoing and has not yet been peer-reviewed — figures are preliminary and will be restated when the paper lands.

Image slot

GPUNeuroMorph system architecture diagram

1600×1000

The Holoscan operator graph — probe input to rendered surface without leaving the GPU.

Image slot

Surface error heat map across the femur

1600×1000

Surface error distribution against the ground-truth donor geometry.

What it came to

The result that mattered was not the accuracy number on its own, it was the accuracy number holding at seven milliseconds. Anything slower than a frame and the surgeon feels the system thinking; anything looser than half a millimetre and the plan is not worth building on. Removing the preoperative CT removes a scan, a wait, and a dose from the patient's path to surgery.